Abstract
There has been increasing interest in using neuroimaging measures to predict psychiatric disorders. However, predictions usually rely on large brain networks and large disorder heterogeneity. Thus, they lack both anatomical and behavioural specificity, preventing the advancement of targeted interventions. Here, we address both challenges. First, using resting-state functional MRI, we parcellated the amygdala, a region implicated in mood disorders, into seven nuclei. Next, a questionnaire factor analysis provided subclinical mental health dimensions frequently altered in anxious-depressive individuals, such as negative emotions and sleep problems. Finally, for each behavioural dimension, we identified the most predictive resting-state functional connectivity between individual amygdala nuclei and highly specific regions of interest such as the dorsal raphe nucleus in the brainstem or medial frontal cortical regions. Connectivity in circumscribed amygdala networks predicted behaviours in an independent dataset. Our results reveal specific relations between mental health dimensions and connectivity in precise subcortical networks.
Introduction
It has become increasingly popular, in recent years, to use measures derived in vivo from human magnetic resonance imaging (MRI) to predict health outcomes, including measures of mental well-being. For example, resting-state functional MRI (rs-fMRI) connectivity measures can predict whether a person suffers from, or will respond to treatment for, Generalized Anxiety Disorder (GAD), Major Depressive Disorder (MDD), and obsessive-compulsive disorders (OCD) 1–4. The prediction accuracies achieved in these types of studies are often impressive and typically reach values between 60-80%. Yet, in the majority of cases, predictions rely on a large number of brain regions, networks or edges. Hence the impressive prediction accuracies come at the significant cost of reduced anatomical specificity.
Despite the critical importance of such studies for diagnosis and prognosis, a lack of anatomical specificity may be problematic when the aim is a mechanistic understanding of the disease to support targeted treatment interventions. Identification and characterization of specific circuits may be necessary for establishing the nature and variants of the illness and it may be critical for developing new treatments that involve manipulation of brain activity in specific circuits.
A second problem is that unsupervised decoding methods, although powerful, are often agnostic to anatomical priors. Yet a large body of evidence has established the roles of specific neurotransmitter systems and particular brain regions in mediating important functions implicated in mental health. Limbic structures that mediate emotional processing and their connections with frontal cortical regions are consistently reported to play an important role. One key hub within this network is the amygdala 5–9. Removal or disruption of this region reduces fear and anxiety responses 10–12. Positron-emission tomography in depressed patients shows abnormal metabolism in amygdala and connected subgenual frontal cortex 6,8,13. And the amygdala is one of the key regions for regulating and expressing emotions 5,10,14,15. An aim in the current study was, therefore, to examine the degree to which it is possible to explain variance in mental well-being across humans, including social and emotional behaviour, in relation to the functional connectivity of identifiable neural circuits – those centred on the amygdala. The monosynaptic connections of the amygdala to specific cortical and subcortical regions have been established for some time in animal models including primates 16 and there is increasing knowledge of the behaviours mediated by amygdala interactions 9,11,17,18.
If, however, a decision is taken to focus on a brain region such as the amygdala then a third problem arises. Many of the key brain areas with which it interacts are in the brainstem where it has been difficult to image activity. Moreover, such regions have very specific connections to particular subnuclei within the amygdala. Therefore, our first step was to parcellate the human amygdala into constituent functional subunits. We took advantage of the high-quality data acquired as part of the human connectome project (HCP; 19). Using resting-state measures from 200 healthy participants, we reliably identified seven amygdala nuclei within each hemisphere. This parcellation was replicated in two additional datasets acquired at 3T (n=200) and 7T (n=98). We also invested considerable effort in developing a refined data pre-processing pathway that focused on the removal of breathing related artefacts that allowed us to examine activity even in brainstem regions, several of which exhibit very specific interactions with particular amygdala subnuclei.
In tandem with improving anatomical specificity, we also aimed to tackle another major problem in relating baseline neural measures to mental well-being. Namely, the disorders themselves are ill-defined and span a broad range of impairments which are not consistently present in all patients diagnosed with the same disorder 20,21 and which are partly overlapping between disorders. This may be another reason why a classifier trained to distinguish a depressed from a non-depressed person is likely to reveal a broad network of regions instead of mapping onto well-defined and anatomically interpretable brain circuits. If we are able to focus on specific rather than broad symptom categories, we may better be able to relate them to specific brain circuits. Because of the sample we examined, mental health varied on a subclinical scale. Nevertheless, we were able to define latent behaviours by applying a factor analysis to a set of questionnaire scores which generated four highly replicable dimensions of mental health. In our final step, we selected the best predictors in terms of the functional connectivity between amygdala nuclei and other brain regions for each of the four behavioural dimensions. We showed that functional connectivity in a select number of specific amygdala connections predicted each mental health dimension in an independent dataset. Our study provides evidence in a large pool of healthy participants that using an anatomically informed approach and a more sensitive characterization of the behavioural phenotypes related to mental well-being, functional connectivity in precise sets of brain connections can predict subclinical dimensional variation in mental health.
Results
In vivo parcellation of the human amygdala into seven nuclei
The amygdala is composed of anatomically distinct nuclei. Our first aim, therefore, was to use rs-fMRI to provide the best possible in vivo parcellation of human amygdala into its nuclei. Previous work in humans in vivo has delineated two or three subdivisions within the amygdala (e.g. basolateral versus centromedial) 22–24. However, given the quality of the HCP data (e.g., improved sequences, 2mm isotropic resolution, 0.7s temporal resolution, ~1h rs-fMRI per person 25), and as a result of the enhanced processing steps we took to remove physiological confound signals, we reasoned that we might reliably identify a more detailed pattern of anatomical organization within the amygdala.
We generated a group connectome using carefully pre-processed rs-fMRI data from a subset of 200 HCP participants. Additional pre-processing focussed on removal of physiological artefacts and improved the temporal signal-to-noise ratio (tSNR) in amygdala and many of its projection targets and sources (see Methods, Supplementary Methods and Extended Data Fig 1, a-b). We did not include all 1206 HCP participants because these additional pre-processing steps required good quality physiological recordings of respiration and cardiac activity which were not available or not of sufficient quality in a considerable number of HCP participants (see Methods and Supplementary Methods for subject inclusion criteria). The resulting group connectome, containing the average functional connectivity between each pair of brain-ordinates, therefore provided high-fidelity connectivity estimates of otherwise difficult to image regions. This is, for example, illustrated by the average amygdalae to whole brain connectivity (Fig 1a), which, in line with previous work 16, highlights overall strong functional connectivity between the amygdala and lateral temporal and temporal pole regions, ventral caudal medial frontal cortex (BA32 and BA25), thalamus, hypothalamus, and ventral striatum. This average amygdala functional connectivity pattern was replicated in two additional HCP datasets (3T: n=200; 7T: n=98; Extended Data Fig 2; see Methods and Supplementary Methods for further details).
Figure 1. Average amygdala functional connectivity and definition of amygdala clusters.
a, Average amygdala connectivity: A group connectome was generated from resting-state fMRI (rs-fMRI) data of n=200 3T young-adult HCP participants using an improved pre-processing pipeline to correct for physiological noise (Extended Data Fig 1). The average functional connectivity of all amygdala voxels to the rest of the brain, corrected for global absolute connectivity strength, shows patterns that would be expected from tracer studies, for example strong connectivity of the amygdalae with subgenual ACC, hypothalamus, and ventral striatum. Scale bar denotes Pearson’s r. This pattern was replicated in two independent datasets containing n=200 additional 3T-HCP participants and n=98 7T-HCP participants (Extended Data Fig 2). b, Amygdala clusters: Hierarchical clustering was performed on the similarities between the whole-brain functional connectivity patterns of different amygdala voxels to identify amygdala subdivisions sharing connectivity profiles. Seven subdivisions were identified (left: horizontal; middle: coronal; right: saggital view), showing strong symmetry across hemispheres and strong resemblance with subdivisions identified from histology and high-resolution post-mortem structural neuroimaging. The parcellations obtained in two independent datasets closely reproduced these nuclei subdivisions (Extended Data Fig 4), but for consistency, the parcellation shown here was used for all analyses reported in this study.
To identify subdivisions within the amygdala, hierarchical clustering was performed on the similarity matrix which summarized the similarities between the functional connectivity of each amygdala voxel and all other points across the whole brain. The similarity matrix was computed based on the group average connectome. It thus ignored variability present across subjects, but instead focused on variability across voxels within the amygdala in terms of their functional connectivity to the rest of the brain that was present across subjects. Hierarchical clustering resulted in parcellations of the amygdalae into increasing numbers of clusters. We evaluated parcellations obtained at different clustering depths, based on anatomical considerations such as cluster location, size, and hemispheric symmetry, and in relation to known subdivisions26–29 (see Methods and Supplementary Methods).. Using these criteria, we chose a parsimonious and anatomically plausible parcellation for further analyses which contained seven subdivisions in each hemisphere (Fig 1b; see Extended Data Fig 3a for shallower and deeper clustering steps). We replicated this parcellation in two additional datasets (3T: n=200; 7T: n=98; Extended Data Fig 4), but used the original parcellation throughout the manuscript. Because this parcellation was obtained from the group connectome and thus the same for each individual, it did not introduce bias in subsequent analyses focusing on individual differences.
Several anatomically plausible features naturally emerged in this parcellation. First, as expected, clusters were nearly symmetrical across left and right hemispheres (Fig 1b). Importantly, this was not the consequence of constraints enforced by the clustering algorithm. Second, the clusters were spatially cohesive but differed in size. For instance, a putative central nucleus contained 30 voxels per hemisphere, but the ventrolateral nucleus contained 50 voxels (for a complete list of cluster sizes, see Supplementary Table 1). Finally, the clusters were located in such a way that a clear progression from ventro-lateral to dorso-medial and from ventral-anterior to dorsal-posterior could be observed, thus corresponding to organizational principles reported previously (Figure 1b; 29).
To facilitate links to other studies, we assigned each cluster a putative label, corresponding to nuclei that have previously been identified (see Methods). As a guide, we used the best match in size and position when comparing our clusters with several atlases of the human amygdala 28–30 (Fig 2a). The seven nuclei were labelled central nucleus (Ce), cortical nuclei (CoN), auxiliary basal or basomedial nucleus (AB/BM), basal nucleus (B), and lateral nuclei (dorsal portion: LaD; intermediate portion: LaI, ventral portion, containing portions of basolateral: LaV/BL). The rationale for our choice of these labels is further explained in the Methods, and its correspondence to nuclei labels in other macaque and human investigations is summarized in Supplementary Table 2. This parcellation and the corresponding labels were used in all analyses reported from this point onwards.
Figure 2. Amygdala nuclei and their profile of functional connectivity to regions of interest.
a, Naming of nuclei: Labels assigned to the seven amygdala subdivisions obtained from hierarchical clustering: Ce = central nucleus, CoN = cortical nuclei, B = basal, AB/BM = auxiliary basal/basomedial, LaV/BL = lateral (ventral part) containing aspects of basolateral, LaI = lateral (intermediate part), LaD = lateral (dorsal part). b, Strength of functional coupling (group average): Average resting-state functional connectivity from the seven nuclei to 28 regions of interest (ROIs) defined a priori based on their known connectivity with the amygdala and potential role in regulating emotions and mental well-being. This highlights strong functional connectivity of subgenual cortex (area 25) to the entire amygdala, but particularly to basal subdivisions, in line with tracer work. Similar profiles are observed for posterior OFC (pOFC) and the subgenual portion of area 32 (s32). By contrast, subcortical and brainstem regions most strongly connect with the central nucleus as expected. The mean functional connectivity between ROIs and nuclei was replicated in two datasets (Extended Data Fig 5). Sale bar denotes Pearson’s r. c, Regions of interest (ROIs): Masks of all ROIs used in this study. For details on their definition, please refer to the Methods. NAc=Nucleus Accumbens; BNST=bed nucleus of the stria terminalis; vl/dPAG=ventrolateral/dorsal periaqueductal grey; SN=substantia nigra; RN_DR/RN_MR=dorsal and median raphe nuclei; LC=locus coeruleus. Definitions of cortical regions were taken from Glasser et al., 2016.
One of the main aims of this study was to identify functional connectivity between amygdala nuclei and other brain regions that helps regulate functions implicated in mental health variation (e.g., sleep and emotion variation). To identify regions of interest (ROIs) with which the amygdala interconnects, we therefore focused on regions central to these processes (Fig 2c) and with known mono- or di-synaptic connectivity with the amygdala. In the brainstem, we defined ROIs in locus coeruleus (LC), dorsal and median raphe nuclei (DRN, MRN), dorsal and ventrolateral periaqueductal grey (dPAG, vlPAG), and substantia nigra (SN)31–35. Subcortically in the forebrain, we included the bed nucleus of the stria terminalis (BNST) and the nucleus accumbens (NAc). In cortex, we focussed on medial areas 24, 25, 32, 9m, posterior OFC, and frontal operculum (FOP) which, on the basis of their similarities with areas in the monkey brain are the frontal areas most likely to be connected with amygdala 36,37. We also considered the prefrontal areas 46 and 9/46 on the lateral surface because stimulating them both affects amygdala threat-related reactivity38. We used ROIs from a recent parcellation 39 which further subdivides area 24 into a24, p24, and a24pr (the most posterior mid-cingulate region p24pr was not included), area 32 into s32, p32, d32, a32pr, and p32pr, frontal operculum into FOP1-5, and area 9/46 into 9-46d, a9-46v, and p9-46v, and which identifies a pOFC region (for more details, see Methods and Fig 2C). For subcortical ROIs, we used established ROIs from published atlases (see Methods) because contrast-based delineation of brainstem nuclei was not available as part of the HCP data. Supplementary Table 1 summarizes all included ROIs with their respective sizes.
Fig 2b shows the average functional connectivity from each of the seven amygdala nuclei, merged across hemispheres, to the above-defined 28 cortical, subcortical and brainstem ROIs. While functional connectivity is strongly influenced by the presence of a monosynaptic connection between areas and plastic changes in those pathways, it also reflects multi-synaptic interactions between regions40. Nevertheless, the pattern of functional connectivity observed for the amygdala nuclei exhibited several features reminiscent of animal tracer studies: all amygdala nuclei had strong functional connectivity with areas in ventral, caudal medial frontal cortex and caudal orbitofrontal cortex, including areas 25, pOFC, and s32 as might be expected from non-human primate studies 16,41–43.
Functional connectivity to these regions was strongest for the basal (B, AB/BM) and cortical nuclei (CoN). The same amygdala nuclei had strong functional connectivity with lateral prefrontal regions (46 and 9/46), but the sign was inverted, suggesting negatively correlated BOLD fluctuations. In stark contrast, the central (Ce) nucleus had the strongest functional connectivity to the majority of subcortical and brainstem regions such as Nac or dPAG. These average functional connectivity patterns between amygdala nuclei and ROIs were replicated in two separate datasets (entire matrix: n=200 3T: two-tailed Pearson’s r(194)=.968, p=8.82e-119, CI = [.958, .976]; n=98 7T: two-tailed Pearson’s r(194)=.884, p=3.92e-66, CI = [.850, .912]; Extended Data Fig 5).
Behaviour: latent dimensions capturing mental well-being
Having established the anatomical plausibility of mean functional connectivity values between amygdala nuclei and our ROIs, given published tracer work, we next sought a characterization of dimensions related to participants’ mental well-being. While the HCP data does not include patients with a clinical diagnosis of a mental health disorder, we sought to examine variance in mental health dimensions along a continuum spanning the subclinical range. We thus selected all behavioural scores available in the HCP data that captured aspects of emotional and psychological well-being, sleep quality, and personality type (see Methods). A total of 33 markers were included which involved measures from the NIH Toolbox ‘Emotion’ (subscales: Psychological well-being; Social relationships; negative affect; stress & self-efficacy), The Pittsburgh Sleep Questionnaire, the Big Five, and the UPenn Emotion Recognition Test. We reasoned that some scores were capturing similar behavioural phenotypes which might have an underlying common cause. To capture such common ‘latent’ dimensions that produce these mental well-being scores, we performed a factor analysis which resulted in four main factors (see Methods; Fig 3a and Extended Data Fig 6).
Figure 3. Latent behavioural dimensions capture distinct aspects of mental well-being.
a, Factor loadings onto behavioural scores: A factor analysis conducted based on 33 behavioural scores (Table 1) available as part of HCP revealed four factors. The loadings for each factor are shown in different colors, corresponding to the four rows. The highest five contributing behavioural scores are shown in order of their contribution (absolute loading) on the right. This shows that the four factors capture quite distinct dimensions of participants’ mental well-being which we summarized as ‘Social and life satisfaction’, ‘Negative emotions’, ‘Sleep’ (problems), ‘Anger and rejection’. The four factors replicated when the factor analysis was performed on all 1206 HCP participants (see Methods), or only the subset of 3T and 7T participants included here (Extended Data Fig 7). b, Correlations between factor loadings (scale bar denotes Pearson’s r).
The first factor emphasized the impact of social support and general life satisfaction, with a strong negative loading onto loneliness and positive loadings onto emotional support, friendship, life satisfaction and purpose (thus, cutting across the sub-scales of ‘psychological well-being’ and ‘social relationships’ within the NIH Toolbox). The second factor, by contrast, loaded strongly onto negative emotions, including fear, stress and sadness (all within the subscale ‘negative affect’ of the NIH Toolbox). The third factor loaded almost exclusively onto sleep-related markers, assessed as part of the Pittsburgh Sleep Questionnaire. It loaded negatively onto the amount of sleep but positively onto sleep troubles such as bad dreams, wakeups, and lack of sleep quality. Finally, the fourth factor loaded onto anger and physical aggression, hostility, and feelings of being rejected, including negative loadings onto agreeableness (Fig 3a and Extended Data Fig 2; Table 1). The factor analysis was replicated in several other datasets (Extended Data Fig 7). Notably, because the factor analysis focuses on behavioural data rather than neural data, it can be employed with all HCP datasets and not just the subset of data with the highest quality physiological recordings of respiration and cardiac activity. When the analysis was repeated on the complete set of n=1206 3T HCP participants for maximal robustness, the resulting factors were highly similar (two-tailed Pearson’s correlation between factor loadings n=200 vs n=1206 3T participants: factor 1: r(31)=.94, p=1.5e-16, CI=[.89,.97]; factor 2: r(31)=.93, p=1e-14, CI=[.86,.96]; factor 3: r(31)=.97, p=9.6e-10, CI= [.93,.98]; factor 4: r(31)=.9, p=1.3e-12, CI=[0.8,0.95]; Extended Data Fig 7).
Table 1. Factor loadings onto mental health scores.
| No | Name | LifeSat | NegEmot | Sleep | Anger/ rejection |
|---|---|---|---|---|---|
| NIH Toolbox Emotion Battery (5-point scale) | |||||
| 1 | Anger Affect | -0.04 | 0.63 | -0.08 | 0.21 |
| 2 | Anger Hostility | -0.28 | 0.34 | -0.12 | 0.20 |
| 3 | Anger Aggression | 0.09 | 0.09 | 0.05 | 0.39 |
| 4 | Fear Affect | -0.17 | 0.82 | 0.01 | -0.16 |
| 5 | Feat Somatic | 0.22 | 0.75 | -0.02 | -0.01 |
| 6 | Sadness | -0.31 | 0.67 | -0.04 | -0.03 |
| 7 | Life Satisfaction | 0.64 | -0.10 | -0.20 | 0.02 |
| 8 | Mean Purpose | 0.54 | -0.10 | -0.12 | 0.09 |
| 9 | Positive Affect | 0.67 | -0.23 | -0.13 | 0.16 |
| 10 | Friendship | 0.67 | 0.06 | 0.07 | 0.02 |
| 11 | Loneliness | -0.51 | 0.15 | 0.00 | 0.26 |
| 12 | Perceived Hostility | -0.03 | -0.11 | -0.05 | 0.91 |
| 13 | Perceived Rejection | -0.41 | -0.14 | 0.00 | 0.66 |
| 14 | Emotional Support | 0.81 | 0.24 | -0.08 | -0.10 |
| 15 | Instrumental Support | 0.57 | 0.11 | -0.05 | -0.01 |
| 16 | Perceived Stress | -0.30 | 0.48 | 0.01 | 0.25 |
| 17 | Self-Efficacy | 0.57 | -0.23 | 0.23 | 0.03 |
| Pittsburgh Sleep Questionnaire (scale from 0-9) | |||||
| 18 | minutes to fall asleep (past month) | -0.12 | -0.12 | 0.74 | -0.09 |
| 19 | hours of sleep per night (past month) | 0.10 | 0.16 | -0.30 | -0.16 |
| 20 | sleep trouble: can’t go to sleep within 30 minutes | -0.10 | -0.04 | 0.71 | -0.02 |
| 21 | sleep trouble: wake-up in middle of night or early morning | 0.08 | 0.05 | 0.55 | 0.04 |
| 22 | sleep trouble: had bad dreams | 0.04 | 0.23 | 0.27 | 0.09 |
| 23 | overall sleep quality | 0.00 | 0.08 | 0.53 | 0.07 |
| 24 | how often taken sleep medicine | 0.10 | 0.19 | 0.35 | -0.12 |
| 25 | how often trouble staying awake during the day | -0.04 | 0.13 | 0.09 | 0.19 |
| 26 | how often trouble keeping up enthusiasm during the day | -0.11 | 0.45 | 0.17 | -0.12 |
| 5-factor model | |||||
| 27 | agreeableness | -0.01 | -0.05 | 0.03 | -0.51 |
| 28 | openness | 0.13 | 0.21 | -0.03 | -0.08 |
| 29 | conscientiousness | 0.05 | -0.45 | 0.15 | -0.03 |
| 30 | neuroticism | -0.35 | 0.55 | 0.01 | 0.02 |
| 31 | extroversion/introversion | 0.47 | -0.09 | 0.05 | 0.03 |
| Penn Emotion Recognition Test | |||||
| 32 | number of correct anger identifications | -0.03 | 0.10 | 0.00 | -0.09 |
| 33 | number of correct fear identifications | -0.09 | 0.00 | -0.02 | -0.13 |
Weights obtained for the 33 scores included in the factor analysis are shown for each of the four factors. These weights were multiplied with an individual’s scores to generate the four behavioural dimensions for each participant.
We used the loadings from the four factors multiplied onto participants’ original 33 scores to construct latent behaviours capturing these four dimensions of participants’ well-being. We summarized them as ‘social and life satisfaction’, ‘negative emotions’, ‘sleep’ and ‘anger & rejection’.
Mental health dimensions and amygdala nuclei connectivity
In the next analysis step, we asked whether the functional connectivity between specific amygdala nuclei and ROIs carried information about mental well-being as captured by the four latent mental health dimensions. Unlike for the amygdala parcellation which used group mean functional connectivity values, here we were interested in interindividual differences in functional connectivity. To improve the reliability of neural signals measured from individual participants, we rejected eight participants with outlier functional connectivity values, corrected for confounding variables, and used robust regressions throughout (for further details, see Methods). All subsequent analyses therefore relied on a total of n=393 3T and n=97 7T participants. In addition, to ensure sufficient spread in mental health dimensions (present in the original n=200 3T participants, but not the n=200 3T replication participants), we pooled the data of all 3T participants for further analysis.
We first established whether relationships between nuclei-specific amygdala functional connectivity and mental health dimensions replicated between the two independent (3T and 7T) datasets. We fitted robust linear regression coefficients to capture the relationship between functional connectivity values in each ‘edge’ (e.g., Ce to NAc) and behavioural dimension (e.g., sleep problems), separately for the 3T and 7T dataset. This resulted in 196 regression coefficients (7 amygdala nuclei x 28 ROIs) for each of four behaviours and two datasets. If amygdala nuclei functional connectivity carries no information about mental health dimensions, then, by chance, the correlation between regression coefficients obtained across behaviours in the 3T versus 7T datasets should be zero. To formally test this, we generated a null distribution by shuffling the subject order of the behavioural scores n=10,000 times while keeping the functional connectivity values unchanged. Indeed, by chance, the across-dataset replication of the pattern of regression coefficients was centred on zero (Fig 4a). The similarity between 3T and 7T regression coefficients in the actual data, however, was significantly greater than chance (one-tailed Pearson’s correlation testing for a positive relationship based on the nonparametric permutation null distribution: r(782)=.26; p=.0313, CI=[.2,.33]; Fig 4a,b), showing that relationships between nuclei-amygdala functional connectivity and mental health dimensions were similar across datasets.
Figure 4. Nuclei-specific amygdala functional connectivity shows consistent relationships with interindividual variation in mental health dimensions.
a, Two types of replication: Relationships between interindividual variation in nuclei-specific amygdala functional connectivity and mental health dimensions were examined in two HCP datasets containing n=393 3T and n=97 non-overlapping 7T participants (following outlier rejection). Despite challenges with neuroimaging signals in subcortical regions, relationships were robust and replicable, as established in two ways: (1) Across-dataset replication (left): the similarity of robust regression coefficients capturing the relationship between resting-state functional connectivity for each ‘edge’ (e.g., Ce to NAc) and each of four mental health dimensions was greater than expected by chance across datasets (null distribution generated using shuffled behavioural scores; n=10,000 iterations). (2) Within-subject replication (right): robust regression coefficients estimated on half of the resting-state data (runs 1+2 versus 3+4, from separate sessions) also showed greater-than-chance similarity (one-sided p-values from nonparametric test using a permutation null distribution). b, Similarity of regression weights across 3T and 7T datasets: Visualization of obtained robust regression coefficients for each edge, mental health dimension (columns) and dataset (rows) illustrates their similarity across cohorts. c, d, For each edge, its contribution rDiff to the across-dataset (c) and within-subject (d) similarity was computed as the difference between the correlation achieved when excluding this edge (195 values) and when including all 196 edges (28 ROIs x 7 nuclei). Visual inspection of rDiff values highlights strong similarities between rDiff values in the two replications (C vs D), clear differences between the four behavioural dimensions, and anatomical specificity – e.g., the importance of cortical connections with B and LaD nuclei for predicting life satisfaction, for connections with NAc, other subcortical regions and medial frontal area p32 for predicting sleep, and functional connectivity with the cortical nuclei (CoN) for predicting anger. e, Out-of-sample 7T prediction using 3T-weights: Regression coefficients estimated from the 3T-participants applied to 7T-functional connectivity values to predict 7T-mental health dimensions showed significant out-of-sample predictions for all mental health dimensions except sleep problems (independent one-sided correlations to assess positive relationships).
To provide a first intuition into the anatomical pathways where functional connectivity values most contributed to the across-dataset pattern replication for each behavioural dimension, we computed the difference in correlation coefficient between 3T and 7T patterns obtained when a given edge was included, versus excluded for computing the correlation (difference in Pearson’s r: rDiff; Fig 4c). Several interesting patterns emerged from visually inspecting the obtained maps (Fig 4c). For LifeSat, the importance of functional connectivity values between frontal regions and B and LaD amygdala nuclei was highlighted. For NegEmot, the functional connectivity with LaD contributed most to the across-dataset similarity. In addition, functional connectivity with subcortical regions LC and NAc, as well as pOFC was important. For Sleep, the most striking features were the importance of subcortical edges (such as with Nac and with several brainstem nuclei including SN, dPAG and LC) and functional connectivity between multiple amygdala nuclei and area p32 in medial frontal cortex. Anger was associated with a broader pattern of contributing pathways but mostly with CoN and LaD. Overall, despite some similarities, the contribution of functional connectivity in specific amygdala pathways clearly differed between the four mental health dimensions. For example, the functional connectivity between medial dorsal area p32pr and the basal nucleus (p32pr-B) contributed to the prediction of life satisfaction (blue; rDiff=.0156) but none of the other three behaviours (all rDiff<.0002), while functional connectivity with Nac and p32 was particularly relevant for sleep but less for life satisfaction, negative emotion or anger.
Having established that the overall relationship between nuclei-specific amygdala functional connectivity and mental health dimensions is similar between datasets, we next asked whether we could predict individual 7T participants’ behavioural scores using regression coefficients estimated from the 3T data. In other words, we examined whether we could predict mental health dimensions in completely held-out data (7T) using a weighting of nuclei-specific amygdala functional connectivity values derived from an independent dataset (3T). For each behavioural dimension, the 196 robust regression coefficients estimated from the 3T data for all nuclei functional connectivity values were applied to the functional connectivity values of individual 7T participants to obtain their predicted behavioural scores. This out-of-sample prediction was significant for life satisfaction, negative emotions, and anger (one-tailed correlation between predicted and true behaviour for the 7T data testing for a positive relationship: lifeSat: r(95)=.187, p=.0335, CI=[-.01,.37]; negEmot: r(95)=.219, p=.0155, CI=[.02,.4]; anger: r(95)=.226, p=.0143, CI=[.02,.4]), but did not reach significance for sleep (r(95)=.05, p=.31, CI=[-.15,.25]; Fig 4e).
Given that medial temporal lobe areas are considered areas of high drop-out and low signal-to-noise, we performed a second replication to further demonstrate the consistency with which nuclei-specific amygdala functional connectivity predicted dimensional variation in mental health. This time, we examined the consistency within-subject rather than across-dataset (pooling across 3T/7T datasets, n=490). We divided the full resting-state data of each participant in two halves (run 1+2 versus 3+4, acquired in separate sessions on different days) and again computed robust regression coefficients to predict the four behavioural dimensions, but this time separately using resting-state functional connectivity values extracted from only the first or second half of the full resting-state data. A null distribution obtained using shuffled behavioural values estimated the similarity of regression coefficients between the two halves (i.e., sessions) expected by chance. The true functional-connectivity-behaviour relationship between the first and second half was significantly greater than expected by chance (one-tailed Pearson’s correlation assessing a positive relationship: r(488)=.47; p=.014, CI=[.42,.52]; Fig 4a). The anatomical pathways where functional connectivity most contributed to this within-subject replication were highly similar to those most contributing to our previous across-dataset replication (Fig 4d). Together, these analyses show that despite substantial noise and difficulties in neuroimaging subcortical regions such as the amygdala44, specific patterns of variation in functional connectivity both within- and between-datasets consistently related to participants’ mental health dimensions.
Characterizing amygdala networks predicting mental health
Having established that the overall pattern of amygdala nuclei functional connectivity carries relevance for mental health, we next explored whether functional connectivity in smaller sets of nuclei-specific edges may be able to predict variation in our four mental health dimensions. So far, all predictions relied on the full set of 196 functional connectivity values from 7 amygdala nuclei to 28 ROIs, but our initial visualization of the pathways that most contributed to similarities both across and within datasets suggested that specific subsets of edges may be particularly important in each case (Fig 4c,d); indeed, it is possible that inclusion of non-contributing predictors may only add noise to the prediction.
To test this in an unbiased way, we iteratively included an increasing number of functional connectivity values from 1 to 196 as predictors, based on their absolute robust linear regression coefficient in the 3T participants, and tested all predictions in the held-out 7T participants. In other words, as before, the 3T-coefficients were applied to 7T-functional connectivity values to generate out-of-sample 7T-predictions, but this time using subsets of increasing numbers of edges, determined based on the order of their contribution in the 3T data. Predictions were evaluated as the correlation between predicted and true behavioural scores in the 7T participants. By definition, adjacent predictions only differ from one another by the inclusion of one additional edge, which necessarily implies interrelationship between predictions. Thus, rather than evaluating each step between 1 and 196 edges, we used this analysis to identify (a) the smallest significant and (b) overall best prediction and examined these statistically in a way that corrected for the number of models (n=196) and behavioural dimensions (n=4). To examine significance in this way, we generated two null distributions with the same procedure but using shuffled behavioural scores (see Methods): (a) for the smallest number of edges to reach significance; and (b) for Pearson’s r at the overall best prediction expected. This showed that our smallest significant networks were smaller on average than expected by chance (lifeSat: n=68, negEmot: n=2, sleep: n=14; anger: n=13; resulting average: n=24.25, one-sided likelihood of this average n value given permutation null distribution: p=.0032). In the same way, the average Pearson’s r at the global peak was higher than expected by chance (lifeSat: r=.197; negEmot: r=.22; sleep: r=.19; anger: r=.24; average: r=.2118, one-sided likelihood of this average r value given permutation null distribution: p=.0068). Importantly, for all four mental health dimensions, the earliest significant and top prediction accuracy was reached before n=196 edges. Fig 5b,c depict the smallest amygdala nuclei networks that reached significance for each behavioural dimension; Extended Data Fig 8 additionally describes the edges associated with the top prediction accuracy.
Figure 5. Functional connectivity in smaller sets of specific amygdala nuclei connections is predictive of interindividual variation in mental health dimensions.
a, Size and nature of amygdala nuclei networks predictive of mental health dimensions: Predictions achieved with subsets of edges between 1 and 196: robust regression coefficients estimated from 3T-participants were applied to 7T-functional connectivity values to predict interindividual differences in mental health dimensions in the held-out 7T data (as done in Fig 4e using all 196 edges). Prediction accuracies are shown as the correlation between true and predicted mental health dimensional scores in the 7T participants, but were only statistically evaluated at the peak (grey arrow: ‘global peak’) and to derive the smallest number of edges that reached a significant out-of-sample prediction (black arrow: ‘smallest sig’; see Methods for generation of null distribution using permutation; for precise p-values in each case, see Results); coloured bars show accuracy when including edges in order of their absolute regression coefficient in the 3T data; black curve indicates performance using the same number of edges but included in random order (n=10,000 shuffles; error bars denote SEM); black line at r=0.168 indicates threshold for significance at p<0.05 purely for visualization (grey line: p<0.1); second row shows the same but zoomed in on the first 70 edges. For all behavioural dimensions, smaller sets of amygdala nuclei functional connectivity values achieve a significant out-of-sample prediction. In general, except for life satisfaction, using the top 3T edges is better than a random selection of the same number of edges. b, c Illustration of the prediction (scatterplot, b) and associated anatomical network (contributing edges shown as fingerprints, c) for the smallest number of edges that achieved a significant out-of-sample prediction (indicated using a black arrow in a). Fingerprints shows ROIs on the circumference (dark=subcortical), amygdala nuclei in the centre (colour-coded); line width denotes the size of the absolute 3T regression coefficient; line style denotes its sign (continuous=positive; dashed=negative).
For life satisfaction, functional connectivity in n=87 connections led to the strongest out-of-sample prediction (r=.197 compared to r=.187 reported above for all n=196 edges; Fig 5a) and the smallest network that reached significance was at n=68 edges (r=.172). The relationship between true and predicted life satisfaction scores achieved for the 7T participants for this number of edges (and based on regression coefficients estimated from 3T participants) is shown in Fig 5b.
To gain insight into the anatomical pathways that carried relevance for life satisfaction, Figure 5c also displays a fingerprint of the functional connectivity values that contributed to the prediction of life satisfaction for n=68 edges (see Extended Data Fig 8 for analogous fingerprints at an earlier peak of n=19 and the top peak at n=87 edges). Fingerprints in Figure 5c also show how strongly (line width) and with which sign (dashed = negative) each edge entered the prediction of life satisfaction scores. The strongest predictor was functional connectivity between LaD and a medial prefrontal region (a24pr).
This and subsequent top edges (see fingerprint for n=19; Extended Data Fig 8) were predominantly between medial frontal and frontal opercular regions, in particular a24pr and p32pr and FOP3 and 4, and the amygdala nuclei LaD (red) and B (light blue). Stronger functional connectivity here was associated with higher life satisfaction scores (positive linear regression coefficient as shown by the continuous line). By contrast, functional connectivity with several subcortical and lateral prefrontal regions (BNST, RN_DR, p9-46v, a9-46v) was predominantly with LaI (yellow) and functional connectivity here contributed negatively (dashed lines) and more weakly (thinner lines compared to cortex) to the prediction of life satisfaction, indicating lower connectivity was associated with higher life satisfaction scores. Note, however, that most of the identified subcortical and cortical connections had small negative average functional connectivity (compare Fig 2). Taking into account both the sign of average functional connectivity and the signs of the correlation coefficients relating functional connectivity with life satisfaction, suggests less pronounced negative functional connectivity between B/LaD and medial cortical regions and more pronounced negative functional connectivity between LaI and subcortical/lateral PFC structures was associated with improved life satisfaction scores.
For negative emotions, peak accuracy was achieved when including functional connectivity in only n=17 edges (r=0.224; Fig 5a), and this prediction was better than that achieved when including all edges (r=0.219). The earliest significant prediction was reached at n=2 edges (r=0.195).
The associated connectivity fingerprint (Fig 5c) showed that, unlike for life satisfaction, the strongest predictors for negative emotion scores were between amygdala nuclei and subcortical regions, in particular nucleus accumbens and locus coeruleus (NAc-B, NAc-AB/BM, LC-Ce; see fingerprints in Fig 5c and Extended Data Fig 8). Stronger functional connectivity in these subcortical pathways was associated with more pronounced negative emotions (positive regression coefficient: continuous line), and was predominantly with superficial amygdala nuclei Ce, CoN and B. At the peak of n=17 edges, only few edges between amygdala nuclei and cortical regions contributed to the prediction of negative emotion scores. These were predominantly with LaD and with more ventral posterior aspects of PFC (area 25, p32, pOFC) and they tended to have negative regression coefficients (dashed lines). The majority of edges relevant for negative emotions showed positive functional connectivity with the amygdala on average, which indicates stronger amygdala connectivity with specific subcortical regions and weaker amygdala connectivity with specific cortical regions was associated with more pronounced negative emotions.
Sleep problems were the only behavioural dimension which could not be predicted based on the full set of edges. However, for sleep, a smaller set of edges performed considerably better. The top prediction for sleep was achieved at n=17 edges (r=0.189), and the earliest significant prediction was achieved at n=14 edges (r=0.185).
The anatomical fingerprints of connections predictive of sleep problems looked quite distinct from those predictive of life satisfaction and negative emotions (Fig 5c). Several strong predictors were between the Ce amygdala nucleus and subcortical/brainstem regions, first dPAG and NAc (Fig 5c) and later also LC (n=17; Extended Data Fig 6). Functional connectivity here tended to be positive on average, and positively predicted sleep scores (continuous line). This suggests stronger Ce-subcortical functional connectivity was associated with more pronounced sleep problems. Cortical connections that helped predict sleep problems were predominantly with medial frontal regions, in particular p32 but also d32 and 9m, and the B nucleus, but not with lateral PFC or FOP areas. The identified edges had negative (dashed) and weaker (thin) regression coefficients and were on average mostly positive, suggesting lower basal nucleus to medial prefrontal functional connectivity related to higher sleep problems.
Anger/rejection reached peak out-of-sample prediction accuracy at n=62 edges (r=0.237), which was better than the prediction achieved with the full set of n=196 edges (r=0.226). The earliest significant prediction was observed at n=13 edges (r=0.174).
The fingerprints illustrating the anatomical pathways predictive of anger/rejection scores were distinct from the other three behavioural dimensions (Fig 5b). One notable feature was that almost all edges entered the prediction with positive regression coefficients (continuous lines), both for functional connectivity with subcortical and cortical regions. Also, most predictive edges were with the cortical nuclei CoN of the amygdala (mid-blue), with LaV/BL (dark red) and LaI (yellow) having the second and third largest number of relevant predictors at the peak prediction (Extended Data Fig 8). Many identified edges had negative or small/modulatory functional connectivity on average, suggesting overall less pronounced negative connectivity in both subcortical and cortical amygdala networks was associated with higher anger/rejection scores.
To test whether, for all four behavioural dimensions, the top predictors in the 3T data, i.e., the edges with highest absolute regression coefficients, indeed performed better at predicting 7T behavioural scores out-of-sample than the same number of randomly picked edges, we repeated the above procedure of including an increasing number of edges from 1 to 196 10,000 times but this time using randomly picked sets of 1 to 196 edges from the full set of 196 amygdala-to-ROI connections. The resulting null distribution is shown in black in Fig 4a. For three out of four dimensions, all except life satisfaction, the top edges selected from the 3T data performed better in predicting the 7T behavioural scores than random subsets of edges of the same size (% of predictions above the null up to 70 edges (2nd row in Fig 5a): lifeSat: 27.14%; negEmot: 98.57%; sleep: 74.29%; anger: 90.00%; % of predictions above the null up to 196 edges (1st row in Fig 5a): lifeSat: 71.94%; negEmot: 69.39%; sleep: 60.71%; anger: 58.67%). This shows that there is consistency between datasets not only in terms of the overall pattern of regression weights, but also in terms of which smaller sets of connections carry particular relevance for predicting a given mental health dimension.
Comparing nuclei-specific and whole-amygdala predictions
In the next step, we tested whether parcellating the amygdala into subnuclei increased the accuracy of predictions of mental health dimensions. We repeated the robust regressions for the amygdala as a whole, i.e., using functional connectivity between the entire amygdala and the same set of 28 ROIs. As before, the robust regression weights of all 28 edges derived from the 3T participants were applied to the 7T participants to obtain out-of-sample predictions. This produced significant predictions of negative emotions (one-sided Pearson’s correlations testing for positive relationships: r(95)=0.2, p=0.026, CI=[0,.38]) and anger (Pearson’s r(95)==0.18, p=0.037, CI=[-.01,.37]; Fig 6a). However, predictions for all four behavioural dimensions were worse when considering whole-amygdala rather than nuclei-specific amygdala functional connectivity (LifeSat: nuclei: r=0.19, whole: r=0.17; NegEmot: nuclei: r=0.22, whole: r=0.20; Sleep: nuclei: r=0.05, whole: r=-0.09; Anger: nuclei: r=0.22, whole: r=0.18; compare Figs 4E and 6A). To account for the number of predictors, we also generated separate null distributions for the nuclei- and whole-amygdala versions to obtain directly comparable p-values, but this did not change the conclusions (see Supplementary Methods and Results). The predictions of all four mental health dimensions still significantly benefited from considering the functional connectivity of separate amygdala nuclei as opposed to the amygdala as a whole (Fig 6b). The same conclusion was reached when comparing the peak accuracy instead of the prediction achieved with the full set of edges (Extended Data Fig 9a and Supplementary Results; fingerprints and scatterplots are shown for the earliest significant peak (Fig 6c,d) and global peak (Extended Data Fig 9b) for comparison with the nuclei version). Thus, despite their small size, considering the functional connectivity of individual subcortical nuclei benefitted predictions of mental health dimensions.
Figure 6. Parcellating the amygdala improves the accuracy for predicting interindividual differences in mental health dimensions.
a, Sensitivity of whole-amygdala as opposed to nuclei-specific amygdala coupling: Out-of-sample predictions achieved when considering the functional connectivity of the whole amygdala with our 28 a priori ROIs, instead of nuclei-specific functional connectivity, are less precise, but still significant for two of the four mental health dimensions (negative emotions + anger; independent one-sided correlations to assess positive relationships). b, Across-dataset: To control for differences in the number of predictors for nuclei-specific and whole-amygdala functional connectivity (28 vs 196), out-of-sample 7T predictions (shown in Fig 4a and 6a) were evaluated based on their own null distribution. Nuclei-specific predictions (top) were still superior to whole-amygdala predictions (bottom) for all four mental health dimensions (coloured bars indicate Pearson’s r overlaid on the null distribution; for statistics see main text); this was also true when looking at peak predictions achieved using smaller sets of edges (Extended Data Fig 9). c,d, As in Fig 5, c illustration of the prediction (scatterplot, c) and associated contributing edges (fingerprint, d) are shown for the smallest set of edges that reached significance (sleep never reached significance; life satisfaction was significant when using fewer than 28 edges) which highlights clear differences between mental health dimension (correlations are shown for illustration using the smallest network that reaches significance in a one-sided correlation, if applicable).
Predicting dimensional and aggregate mental health indices
So far, we have assumed that dimensional markers of mental well-being that allow us to capture functionally meaningful processes (e.g., those related to negative emotions or sleep) are more likely to relate to amygdala functional connectivity than more general markers of mental health such as an overall depression score. To test this assumption, we repeated our analyses using the DSM score and the score of the ASR subscale AnxD. Out-of-sample predictions using all 196 nuclei-specific 3T-weights applied to 7T-participants’ functional connectivity values were not significant for either score (DSM: r(95)=0.11, p=0.13, CI=[-.1,.29]; ASR AnxD: r(95)=0.1, p=0.13, CI=[-.09,.31]; Fig 7a and Extended Data Fig 10). This suggests that amygdala nuclei functional connectivity values are more useful for predicting mental health dimensions as opposed to broad markers like depression. Out-of-sample predictions of three of our dimensional behaviours, all except sleep, were better than the predictions achieved for DSM or ASR-AnxD (Fig 7b,c). Thus, understanding the contributions of small subcortical networks to mental health benefits from considering functionally meaningful behavioural dimensions, rather than aggregate scores that capture complex sets of symptoms.
Figure 7. Amygdala functional connectivity relates better to dimensional behaviours than overall depression scores.
a, b Predicting depression score instead of dimensional markers: Out-of-sample prediction of 7T participant’s ASR_AnxD scores, using nuclei-specific functional connectivity and regression weights estimated from 3T participants as in Fig 4e, is (a) not significant (for DSM, see Extended Data Fig 10) and (b) less accurate than three out of four of our dimensional behaviours (independent one-sided correlations to assess positive relationships). c, Overall predictions are worse for ASR compared to dimensional scores when using smaller sets of edges (bars shown in Fig 5a for dimensional behaviours are overlaid as coloured lines for comparison); plotting conventions as in Fig5.
Discussion
The need to better describe the biological underpinnings of psychological illness and dimensional variation linked to psychological illness has long been recognized but recently re-emphasized20. Here, we used resting-state fMRI, a common in vivo tool for estimating human brain connectivity, but applied a fundamentally different rationale and approach to the analyses of both neural and behavioural data. Turning first to behavioural data analysis, rather than stratifying a disease such as depression into several biologically meaningful sub-groups 45 or classifying people into categories (e.g. patient vs control), we aimed to define biologically meaningful behavioural dimensions that capture central aspects of mental health that exhibit variation even in the subclinical range. In the neural analysis we were able to predict these mental health dimensions using functional connectivity in a select number of anatomically motivated brain connections. All of this was done using several independent data sets, ensuring robustness and replicability.
We identified four behavioural dimensions which we believe capture distinct aspects of people’s mental health: social/general life satisfaction, negative emotions, sleep problems, and problems with anger/rejection. Rather than using a summary measure, such as e.g. the total depression score, we reasoned that because specific brain connections carry specific combinations of input and output, mappings of behaviour onto the functional connectivity in precise connections are more likely achieved for functionally meaningful behavioural units. We obtained such behavioural dimensions using a factor analysis 46,47. In order to link the behavioural dimensions to precise brain connections, we focused on the amygdala. First, we demonstrated that it was possible to identify in vivo seven component amygdala subregions that corresponded to amygdala nuclei. They reliably varied in their functional connectivity in comparison to one another, but they were topological arranged in a similar manner in both hemispheres. This amygdala parcellation replicated across several data sets. Second, we demonstrated that the average patterns of functional connectivity – correlations in the BOLD signals – between each amygdala nucleus and 28 cortical, forebrain subcortical, and brainstem regions were approximately as predicted from anatomical tracer studies. These ROIs were selected a priori based on strong, often monosynaptic, connectivity with the amygdala in animal tracer work and their relevance for mental health related functional processes. We were then able to proceed to the final stage of the study and show that variation in functional connectivity between specific amygdala nuclei and these other regions were predictive of variation in the four mental health dimensions.
Three aspects of our data underlined the importance of the functional connectivity of specific amygdala nuclei. First, for three of our behavioural dimensions, functional connectivity in a small number of less than 15 connections was sufficient to allow prediction of behavioural scores in an independent dataset (Fig 5b and Extended Data Fig 8) and we would expect prediction accuracies to be even greater in a clinical population that includes the extremes of the behavioural distribution. In the context of neuroimaging, our sample size of 490 participants across datasets can be considered fairly large. We therefore believe the reported effect sizes are meaningful. Despite the importance of large network approaches 1–4, an advantage of the current approach is that it provides specific regions and connections as targets for therapeutic intervention involving a range of approaches such as pharmacological, neurostimulation, neurofeedback, or cognitive interventions. Second, variations in nuclei-specific amygdala functional connectivity were associated with better predictions of the mental health dimensions than was possible when just the connectivity of the amygdala as a whole was considered (Figure 6 and Extended Data Fig 9). Finally, in a third test, we established that amygdala nuclei functional connectivity was better suited to predict behavioural dimensions than aggregate depression scores, which was true for three out of four mental health dimensions (Figure 7 and Extended Data Fig 10). Importantly, predictive connections largely differed between the four dimensions of mental well-being (see fingerprints in Fig 5) and only few edges were shared.
We had a strong anatomical prior not only on the importance of the amygdala but the importance of the amygdala’s interactions with specific cortical, forebrain subcortical, midbrain, and brainstem regions thanks to the large body of studies in animal models that have examined these circuits 8–11. As a result of careful fMRI data preprocessing we were able to examine activity not just within medial temporal lobe but even in specific brainstem regions and relate functional connectivity to variation in our indices of mental health. We believe there are other prime anatomical hubs such as ventromedial and subgenual frontal areas that would be worth investigating with a similar approach. It is unlikely that a single region or network is sufficient to fully predict all aspects of someone’s mental well-being 45,48. Nevertheless, we believe it is important to recognize that functional connectivity in individual and identifiable edges may have particular importance. This is a view taken more commonly when considering targeted interventions in mental health such as, for example, using invasive deep brain stimulation which has in some cases led to remarkable improvements in mood 49–51, but which may work particularly well when the right connections between subcortical and cortical regions are targeted 52. Similarly, other non-invasive stimulation approaches such as repetitive transcranial magnetic stimulation (rTMS) are more likely to be successful when targeting or altering specific subcortical circuits38,53–56. Such interventions could become more feasible with advances in non-invasive ultrasound methods 57–60. If variation in these functional connectivity patterns explains variation in behaviour in patients with clinical levels of depression, then our findings suggest interventions targeted at particular nuclei might benefit someone predominantly suffering from sleep problems while targeting others might benefit someone who experiences strong negative emotions.
Our parcellation of the amygdala into seven nuclei strikingly resembled previous amygdala investigations but which were possible only post mortem 28,29. Saygin et al., for instance, scanned at a resolution of 100-150um at 7T and identified nine nuclei which resembled in their size, position and transition patterns the seven nuclei identified here. Previous parcellations based on in vivo data have identified fewer subdivisions 22–24 but the borders identified in those studies still resembled a subset of the borders we identified here thereby underlining consistency in results. The finer grained parcellation we obtained reflected improved image quality and preprocessing pipelines that better controlled for physiological noise. We believe more data with similar resolution and SNR is likely to become available, which will make our method more broadly applicable to the neuroscientific community and to the study of other subcortical brain structures, including in clinical populations. In our data we show that detailed amygdala parcellation is important for achieving the behavioural prediction accuracies reported here (Fig 5). This is unsurprising given known anatomical and functional differences between the amygdala’s small nuclei 16.
The anatomical features of the amygdala networks identified for the different latent behaviours seem plausible in the context of previous work, and consistent across two types of replications (across-dataset and within-subject; Fig 4). We note that, importantly, both feature selection – which determined the anatomical networks to focus on – and estimation of regression weights was performed in an initial dataset (n=393 3T participants) and all predictions were generated out-of-sample (n=97 7T participants).
Social and life satisfaction was most strongly predicted by functional connectivity between the amygdala and regions primarily located in medial and lateral frontal cortex, such as areas a24pr, p32pr, a32pr and d32 as well as several frontal opercular (FOP3, FOP4) and lateral prefrontal areas 39, with less pronounced negative connectivity between these areas and the basal or dorsal lateral amygdala nucleus predicting improved life satisfaction (Fig 4c,d; Fig 5c; Extended Data Fig 8). These areas in or close to the dorsal anterior cingulate cortex (dACC) as well as frontal opercular/insula regions have been linked to aspects of behavioural change and adaption, abilities compromised in anxiety, and are important for arbitrating between exploration and exploitation, a process changed in depression61–64. Even though there is probably little direct connectivity between dorsolateral prefrontal cortex (dlPFC) and amygdala, dlPFC is a stimulation target in depression, and alters amygdala threat responses 38,56. It thus seems unsurprising that functional connectivity between these medial and lateral frontal regions and the amygdala might contribute to overall social and life satisfaction. Medial prefrontal, frontal opercular and lateral prefrontal neurons express receptors for several neurotransmitters, including serotonin (e.g., 5HT-1A, 5-HT-2), noradrenaline (alpha1) and dopamine (D5)65–69. Receptors for these neurotransmitter systems are also expressed in the amygdala, with varying densities across nuclei65. Given the modality of the HCP data used here, it is difficult to establish which neurotransmitter systems are most likely to be mediating the observed effects. However, linking functional connectivity changes with changes in specific neurotransmitter pathways would be an interesting avenue for future research.
There is evidence that resting-state correlations reflect anatomical connectivity40. However, it is worth noting that, while rs-fMRI was used as a proxy for anatomical connectivity here, the patterns in resting-state functional connectivity we identify do not necessarily correspond to monosynaptic connections.
The associations between negative emotions, our second latent behaviour, and amygdala functional connectivity can also be understood in the context of the functions of these areas even if, once again, some of the critical pathways may be indirect. Amygdala functional connectivity with areas pOFC, 25 and several subcortical structures (most prominently LC, NAc) seem plausible. In cortex, weaker connectivity between pOFC or FOP3/4 and LaD, but stronger connectivity between area 25 and B are related to more pronounced negative emotions (the inverse of what was observed for life satisfaction; Extended Data Fig 8). Although pOFC has received little attention, bipolar patients demonstrate reduced grey matter in pOFC 70 and both pOFC and amygdala have been linked to the most basic aspects of stimulus-reward association learning 71. Area 25 has been linked to autonomic and affective regulation and, just like the amygdala, exhibits abnormal metabolism in depressed patients 13. Stimulation of this region or its interconnections may reduce depression 49,51,52. The fourth latent measure of mental well-being we identified, anger and rejection, was also linked to areas 25 and pOFC, in addition to other frontal opercular regions that have recently been linked to the balancing of the most recent outcomes with the wider, more long-term experience of reward 72. However, here functional connectivity with the cortical nuclei of the amygdala, CoN, was of greatest importance (Fig 4c,d; Fig 5; Extended Data Fig 8).
In contrast to cortical regions, stronger rather than weaker subcortical functional connectivity with amygdala nuclei predicted negative emotions. This suggests that diminished cortical-amygdala interaction is accompanied with increased amygdala interaction with subcortical areas linked to the origins of widely branching neuromodulatory systems such as serotonin and noradrenaline (RN_DR, LC) and key targets of other systems such as dopamine (NAc). Noradrenaline mediates stress and stress-related responses and stress-induced dysregulation of the NA system may contribute to the pathogenesis of depression 73. Increasing NA can also be effective as an antidepressant treatment. LC and NAc functional connectivity was also predictive of negative emotions when considering functional connectivity with the whole amygdala (Fig 6d and Extended Data Fig 9). This suggests that a more global functional connectivity pattern between LC/NAc and amygdala may help regulate mood.
The third latent measure of mental well-being captured sleep problems and was linked to a distinctly different connectional fingerprint (Fig 4c,d; Fig 5c). Unlike the other three behaviours, it was not well predicted using the whole set of nuclei-specific functional connectivity values (Fig 4e), but fared better using a smaller number of edges (Fig 5), which comprised functional connectivity in a predominantly subcortical network, in particular with NAc, but no FOP or lateral frontal areas. The NAc is an important projection target of VTA dopamine neurons, and dysfunction of the striatum has been associated with sleep disturbances, with neurons in NAc core particularly important for controlling slow-wave sleep 74.
We note several limitations of our approach: we relied on relatively large data sets (one hour of resting-state per participant) from highly optimized pulse sequences, which may not be available regularly in patients. Some of the results we obtained may reflect the improved image quality and preprocessing pipelines that controlled for physiological noise. However, acquiring cardiac and respiratory traces is easy and uses standard equipment available in most MR facilities and so our approach might be used elsewhere. The HCP data does not include patients with a clinical diagnosis of a mental health disorder and prediction accuracies may be greater in a clinical population that includes the extremes of the behavioural distribution. Finally, we do not compare the amygdala networks described here to networks elsewhere in the brain. It is therefore not possible to conclude that they are necessarily the best possible network of its size for predicting the mental health dimensions studied here. Instead, our data emphasize (a) the value of parcellating subcortical structures to match the nuclei scale at which these circuits are organized when studying functional connectivity; (b) the importance of using functionally interpretable behavioural dimensions, rather than complex aggregate scores; and (c) that small subcortical networks can carry meaningful relevance for predicting mental health dimensions.
In summary, our work suggests that strong anatomical priors derived from animal studies, in combination with neuroimaging data of sufficient quality and resolution, make it possible to forge links between dimensions of mental health and specific neural circuits. Crucially such predictions are facilitated by the identification of mental health dimensions which, even if in the subclinical range, are naturally emerging functional groupings that are more likely to map onto the brain’s functional organization.
Methods
Participants
Data and ethics were provided by the Human Connectome Project (HCP), WU-Minn Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research; and by the McDonnell Center for Systems Neuroscience at Washington University. All participants gave informed consent and were reimbursed for their time ($450 for 3T MRI + interview, $400 for 7T MRI) and travel. Several HCP resting-state fMRI datasets were included for analysis25,75. First, an initial dataset comprised a subset of n=200 out of the full set of n=1206 3T subjects from the HCP young adults data set (n=200; mean age 29 ± .26; age range 22-36; 108 females, 92 males). These were chosen from the full HCP data set (https://www.humanconnectome.org/) based on two criteria: the quality of the physiological variables acquired (both cardiac and respiratory; inspected visually and using summary measures such as their variance over time) and their total DSM/ASR (DSM_Depr, ASR_Totp) score to allow us to maximise subclinical variance across participants (resulting mean DSM score: 4.46, variance: 16.64; mean of all 1206 HCP participants: 4.25; variance: 12.24; mean/variance total ASR score n=200: 37.91, 669.08; n=1206: 37.43, 523.82). For more details on participant selection, see Supplementary Methods. A second dataset of the same size as the first (n=200) was selected for replication from the remaining 3T-participants with complete data (for more details, see Supplementary Methods). The demographics in this second dataset of n=200 3T participants were mean age 28 ± .28; age range 22-36; 99 females, 101 males. A third dataset contained all 7T-HCP young adult participants not already included in either of the 3T datasets and with full resting-state and behavioural data, which left us with n=98 7T-HCP participants (mean age 29 ± .33; age range 23-36; 59 females, 39 males; DSM mean/variance: 3.43, 5.73; ASR mean/variance: 31.79, 253.43; Supplementary Table 3).
Data and minimal pre-processing
Four resting state runs were acquired on a 3T Siemens Skyra or 7T Siemens Magnetom scanner using custom pulse sequences (for details see 19,25). In brief, 3T resting-state runs lasted 14.4 minutes, had a TR of 720ms, TE of 33ms, isotropic resolution of 2mm, 72 slices, and a multiband factor of 8 resulting in 1200 timepoints. 7T resting-state runs lasted 16 minutes, with a TR of 1 s, TE of 22.2 ms, isotropic resolution of 1.6mm, 85 slices, a multiband factor of 5 and in-plane acceleration factor (iPAT) of 2. Two runs were acquired using right-left phase encoding and two using left-right phase-encoding (or anterior-posterior versus posterior-anterior for 7T). Spin-echo images and T1-weighted images were acquired for distortion correction and registration (for more details see 76). We used all four runs of each subject and downloaded the minimally pre-processed HCP data which is described in detail in 25. In brief, these data are distortion-corrected, temporally filtered, projected on to a surface reconstruction obtained from the T1-weighted image while maintaining subcortical voxels (cifti format), and minimally smoothed. These pre-processing pipelines are based on tools from FSL, FreeSurfer (version 5.2) and Connectome Workbench command-line functions. 7T data were subsampled to 2mm to match the resolution of the 3T data. Registration across participants was achieved using multi-modal areal-feature-based surface registration (MSMall) 39.
Additional pre-processing of 3T data
Despite its high quality, there are ongoing discussions on the relatively weak signal in subcortex, compared to cortex in the 3T HCP-YA data. A large majority of published work using these data has consequently focused on cortical regions. Because noise caused by physiological artefacts (e.g. breathing, pulse) is particularly pronounced in brainstem and temporal lobe structures, all key areas for this study, we performed additional corrections for physiological noise in the data. Removal of artefacts caused by physiological signals is not currently incorporated in standard HCP pipelines. We used the PNM toolbox (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/PNM; 77) to generate physiological regressors (a total of 33 regressors comprised of: cosine and sine of basic cardiac and respiratory regressors modelled with an order of 4, and thus 16 regressors; multiplicative cardiac and respiratory terms cos(c+r), sin(c+r), cos(c-r), sin(c-r), each modelled using an order of two, and thus again 16 regressors; plus respiration volume per time (RVT)77). In addition to physiological regressors, we constructed 24 motion regressors from the six motion regressors provided (in the HCP data release, these are stored in Movement_Regressors.txt) (e.g., 76): the six original regressors, their derivatives, and the square of the resulting twelve regressors. We also used independent component analysis (ICA)-denoising as provided with the ‘fixextended’ HCP dataset (melodix_mix and Noise.txt). The motion, physiological and ICA noise regressors were normalized, high-pass filtered and detrended to mimic the pre-processing performed on the data. Then, motion and physiological confounds were aggressively regressed out of the data and ICA components (thus entirely removing any variance explained by physiological or motion parameters), and the noise ICA components were subsequently removed from the data using a soft regression (thus removing only the variance unique to the ICA noise components). See Supplementary Methods for a validation of the additional pre-processing to correct for physiological noise.
The data were demeaned, the variance of the noise in the data normalized (as in 39) and the four runs of each participant were concatenated. This yielded the fully pre-processed data for each participant which contained a total of 4800 time points from the combined 1200 time points of the four resting-state runs for 3T data (n=400 participants), or a total of 3600 time points from the combined 900 time points of the four resting-state runs for 7T data (n=98 participants). Additional smoothing was applied to the surface only for the creation of the group connectome (see below and Figs 1+2; sigma=5mm; no additional smoothing was applied to subcortical structures, including the amygdala). Analyses conducted on individual time series (Figs 4 onwards) did not apply any further smoothing.
Group dense connectome
A group average timeseries was generated from the 200 individual data sets of the first 3T dataset (n=200) using the algorithm ‘MIGP’ 78 (see reference for code and Supplementary Methods for further detail). A dense connectome was created from the average time series using the function cifti-correlation (using Fisher’s z). Ringing artefacts were corrected using Wishart RollOff 39. The same procedure was repeated for the 3T and 7T replication data sets (n=200 and n=98).
Clustering
The full dense connectome was restricted to contain the functional connectivity of voxels in both amygdalae to the rest of the brain (647 voxels x 91282 brain-ordinates). Functional connectivity values were transformed into absolute values (i.e., unsigned ‘strength’ of functional resting-state connectivity). This had the desired effect that the similarity matrix created in the next step was driven by differences in high vs low functional connectivity values rather than differences in large positive vs large negative functional connectivity values. A similarity matrix was computed based on this absolute functional connectivity using Pearson’s correlation coefficient (FSLnets function nets_netmats, part of FSLnets: https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FSLNets). In other words, the similarity matrix captured, for any pair of amygdala voxels, the similarity of their connectivity profile to the rest of the brain. The similarity matrix was fed into a hierarchical clustering algorithm (function nets_hierarchy.m part of FSLnets). We thus obtained clustering solutions of the amygdalae at different depths based on the similarity of different amygdala voxel’s connectivity to the rest of the brain.
To determine the appropriate depth of clustering, we inspected the resulting parcellations for symmetry across hemispheres, a balance between simplicity and detail, and anatomical plausibility. Further details on how we chose the final clustering depth can be found in the Supplementary Methods. Throughout the results, we focused on the cluster solution from depth 12, which when merging corresponding clusters in both hemispheres yielded seven clusters (Figs 1b and 2a). Other shallower and deeper clustering depths are shown in Extended Data Fig 3.
Following the exact same procedure, we closely replicated this parcellation in two additional datasets (3T: n=200; 7T: n=98; Extended Data Fig 4; Supplementary Methods). We used the parcellation generated from the first 3T dataset for all further analyses. Importantly, since this parcellation was obtained from the group connectome, rather than for each subject individually, it did not introduce bias in subsequent analyses focusing on individual differences.
Naming of clusters
The labelling of clusters was largely based on the Atlas of the Human Brain 28 and a post-mortem parcellation at 7T with 100-150um resolution 29. Similarities between our labelling and the labels of those two atlases as well as two common nomenclatures used in non-human primates are shown in Supplementary Table 2. Supplementary Table 1 also summarizes the size of all clusters. For a detailed description of our choice of nuclei labels, see Supplementary Methods, Figs 1B and 2A. Overall, there was good correspondence between the position and labelling of our nuclei, and those reported in 26–29 (Supplementary Table 2).
ROI selection
We had several a priori regions of interest which were informed by prior work, including anatomical work using tracers in macaque monkeys as well as work in humans with mental health disorders. Our ROI selection was thus entirely hypothesis-driven and followed two key criteria: (1) knowledge about a region’s mono- or disynaptic connectivity with the amygdala and (2) knowledge about a region’s importance in mental health/mood disorders. Most of our ROIs fulfilled both criteria. The majority of the ROIs we included are known to have monosynaptic connections with the amygdala (or in the case of some brainstem and dlPFC regions, di-synaptic connections that are made via the hypothalamus or medial PFC, respectively), and most regions with strong amygdala connectivity were included. We left out some regions with monosynaptic connections to the amygdala – for example anterior temporal lobe regions - because we did not have strong reasons to believe they were relevant for the mental health-related behaviours of interest in this study. Other notable omissions are the hypothalamus and hippocampus which we would have needed to parcellate and which it would not have made sense to include as whole structures. All our ROIs are illustrated in Fig 2c. We included a total of eight subcortical and brainstem regions. Please refer to the Supplementary Methods for how the binary masks for these ROIs were created and for a detailed motivation for each one of them. A full list of all ROIs detailing their respective sizes in voxels/vertices is shown in Supplementary Table 1; apart from the median raphe (RN_MR), all ROIs had at least 20 voxels/vertices. ROI selection was performed prior to and independent of subsequent analyses focusing on individual differences.
For analyses of group and individual data, we computed the average time course from each ROI. Thus, while a smaller ROI might have a less reliable average time course compared to a larger ROI, all ROIs were given equal weight in subsequent analyses. In general, we did not find that most effects pertained to larger ROIs, suggesting that even smaller ROIs were of sufficient size to capture meaningful signal. This notion was corroborated by the mean pattern of functional resting-state connectivity between amygdala nuclei and ROIs shown in Fig 2b which we robustly replicated in two independent data sets (Extended Data Fig 5).
Selection of behavioural scores
Instead of using psychiatric scores (e.g., the total depression score), our goal was to define underlying variation in emotional and social wellbeing in the subclinical range but, in particular, in those dimensions of emotional and social wellbeing that might be affected in anxious or depressed individuals. We went through all restricted and unrestricted behavioural markers acquired as part of HCP and selected those that related to aspects of mental well-being. We included 33 behaviours composed of
-
(1)
17 measures from the NIH Toolbox Emotion Battery (www.nihtoolbox.org) 79,80 (a) Six measures of negative affect: Anger Affect, Anger Hostility, Anger Aggression, Fear Somatic, Sadness; (b) three from the Psychological Well-Being toolbox (Life Satisfaction, Mean Purpose, Positive Affect); six from the Social Relationships toolbox (Friendship, Loneliness, Perceived Hostility, Perceived Rejection, Emotional Support, Instrumental Support); two from the Stress and Self Efficacy toolbox (Perceived Stress, Self-Efficacy)
-
(2)
9 measures from the Pittsburgh Sleep Questionnaire 81: minutes to fall asleep (past month); hours of sleep per night (past month); sleep trouble: can’t go to sleep within 30 minutes; sleep trouble: wake-up in middle of night or early morning; sleep trouble: had bad dreams; overall sleep quality; how often taken sleep medicine; how often trouble staying awake during the day; how often trouble keeping up enthusiasm during the day.
-
(3)
5 measures from the five-factor model 82: neuroticism, extroversion/introversion, agreeableness, openness, and conscientiousness83.
-
(4)
Measures from the Penn Emotion Recognition Test: the number of Correct Anger Identifications (ER40ANG) and the number of Correct Fear Identifications.
For details on the number of items and scoring associated with these items, see Supplementary Methods.
Factor analysis and creation of latent behaviours
We conducted a factor analysis on these 33 behavioural markers (z-scored) using Matlab’s function ‘factoran’, with a ‘promax’ rotation. A Scree test based on the first n=100 participants suggested four factors (nFactors package in R with function nScree 84), all of which seemed interpretable upon inspection of their weights. We therefore fixed the number of factors to four. Importantly, the same four factors replicated in our full first dataset of n=200 3T-HCP; see Supplementary Methods for multiple replications of the resulting factor analysis.
The weights obtained for the four factors were multiplied onto the original 33 behavioural markers (z-scored) to construct four summary or latent behaviours per participant. These were summarized and are referred to throughout as ‘social and life satisfaction’, ‘negative emotions’, ‘sleep’ and ‘anger & rejection’.
Mental health relationships with amygdala nuclei coupling
A regression approach with two types of out-of-sample replication was used to identify whether amygdala nuclei functional connectivity was predictive of the four mental health dimensions. To generate regression weights, the data to be predicted, y, was a 393x1 or a 97x1 vector describing the true behavioural score for each 3T or 7T participant. The matrix of potential predictors X was a matrix with 393 x 196 (or 97 x 196) resting-state functional connectivity (FC) values for each participant and the 196 measures of functional connectivity between the areas described above (7 amygdala nuclei x 28 ROIs) which we refer to as “edges”. Such functional connectivity measures are known to be indices of anatomical connectivity, although the connections are not always monosynaptic)40. Outlier participants from the original pool of 400 3T and 98 (non-overlapping) 7T participants were conservatively rejected based on their individual FC values if more than 10% of their FC values across all edges deviated more than 3.5 standard deviations from the mean across participants. This identified seven 3T and one 7T participants as outliers and all analyses were performed on the remaining 393 and 97 participants. Next, confounds were regressed out of the data (see 85), and this was done separately in both the 3T and 7T data. Confounds included (1) a summary statistic quantifying average head motion; (2) weight; (3) height; (4) blood pressure – systolic; (5) blood pressure – diastolic; (6) haemoglobin A1C in blood; (7) cube-root of total brain volume; (8) cube-root of total intracranial volume. A total of 8 confounds were thus regressed out of the matrix X. Both y and X were z-scored.
For generating the plots in Fig 4b, we estimated robust linear regression models (Matlab’s function robustfit) for functional connectivity in each of the 196 edges, four behavioural dimensions, and in the 3T and 7T cohorts. In each case, the resulting robust regression weight captured the relationship between FC and behaviour. Including all edges in one large regression model was not feasible due to the large number of regressors and existing correlations between them.
To test whether the obtained robust regression weights captured meaningful relationships between the functional connectivity of individual amygdala nuclei and the four mental health dimensions, we tested whether regression weights were (a) similar between the 3T and 7T datasets (‘across-dataset replication’) and (b) similar between two halves (corresponding to MR sessions) of the experiment (‘within-subject replication’; Figure 4a). First, for the across-dataset replication, we computed Pearson’s correlation coefficient between the overall pattern of regression weights obtained for the 3T and 7T data (row 1 versus 2 in Figure 4b). To establish whether the obtained correlation was better than predicted by chance, given the level of noise present in brain connections with the amygdala and given our number of connections, we generated a null distribution (Fig 4a) by shuffling the vectors y containing the behavioural dimension n=10,000 times and recomputing the correlation coefficient between the overall pattern of regression coefficients.
Second, to attempt a within-subject replication, functional connectivity values were extracted from half of the resting-state data, separately from either run 1+2 or run 3+4 (which were acquired in two separate sessions on separate days). Robust regression weights capturing the relationship between FC (X) and behavioural scores (y) were then computed separately for the FC values obtained from runs 1+2 versus 3+4. Here we used the merged data from all n=490 3T+7T participants. The similarity between the overall pattern of robust regression weights obtained for the two experimental halves was computed using Pearson’s correlation. Again, to test whether their similarity was greater than expected by chance, a null distribution was generated by repeating this procedure n=10,000 times using shuffled behavioural scores (Figure 4A).
To visualize the contribution of individual FC values to the two types of replications, we computed the correlation between the pattern of robust regression coefficients for each behavioural dimension and compared the correlation coefficient obtained using all FC values with the one obtained when a single FC value, or edge, was left out (which was repeated once for each of the 196 edges). The difference between the full and the left-out correlation coefficient captures how much functional connectivity in an individual edge contributes to the replication (rDiff). Figures 4C+D show rDiff for the two types of replications, respectively, i.e., comparing the 3T and 7T regression coefficients (across-data-set replication) and comparing those obtained from the two experimental halves (within-subject replication).
Knowing that relationships between amygdala functional connectivity and mental health dimensions are consistent within and across datasets does not necessarily imply that coefficients estimated from one dataset will be useful in predicting behavioural scores in an independent cohort. To test if this was possible, we applied the complete set of 196 robust regression weights estimated from the 3T participants’ data to the FC values measured in the 7T cohort, to predict interindividual differences in the 7T-participants’ behavioural scores (Figure 4E). The goodness-of-fit of this prediction was determined as the correlation (Pearson’s r) between true mental health dimensional score and the out-of-sample model-predicted mental health dimensional score.
Throughout the manuscript, p-values for replication analyses (Figs 4A, 6B, 7B) and analyses probing whether FC values could predict behavioural dimensions (Figs 4E, 5B, 6A, 7A) were calculated to address the question of whether there was a positive relationship between two patterns of robust regression coefficients (replication) or between predicted and true behavioural scores (out-of-sample prediction), respectively, and were accordingly one-tailed.
Characterizing amygdala networks predicting mental health
The above predictions were all generated using the full set of 196 FC values. To check whether significant out-of-sample predictions were possible using smaller sets of FC values, we repeated the prediction of 7T-behavioural scores, but this time we iteratively included an increasing number of edges from 1 to 196. The order in which edges were included was based on their absolute regression coefficient estimated from the 3T participants. As above, the selected 3T-coefficients were applied to the corresponding 7T-FC values for each subset of connections to predict 7T behavioural scores, and this procedure was repeated using an increasing number of FC values (from 1 to 196, in steps of 1). Thus, all predictions were again generated out-of-sample and the selection of edges was based on an independent dataset and thus unbiased. Figure 5A shows the goodness of fit (correlation between true and predicted 7T behavioural scores) for 1 to 196 edges. For each behavioural dimension, the obtained correlation coefficients were used to identify (a) the smallest number of edges that led to a significant out-of-sample prediction (r>0.168); and (b) the best possible out-of-sample prediction (maximum r-value). For visualization, Fig 5a shows lines to indicate trend-wise and p<0.05 significance levels, which correspond to Pearson’s r-values of 0.1312 and 0.168 (given n=97, one-sided).
To establish whether (a) the size of the smallest significant network and (b) Pearson’s r at the peak prediction across all 196 models were significant, we generated two permutation null distributions (see Supplementary Methods).
To illustrate the quality of the prediction and the contributing edges, scatterplots and fingerprints were generated for the smallest number of edges that reached significance (by definition, these will just fall under p<0.05; Fig 5b), and for the number of edges that showed the overall best prediction (Extended Data Fig 8). Fingerprints show the size of the 3T regression coefficient for each edge as the width of the line, its sign for predicting the behavioural dimension as the line style (continuous=positive; dashed=negative) and the colour reflects the amygdala nucleus. In Extended Data Fig 8, in addition, a surrounding line indicates whether functional connectivity in the same edge was on average positive, negative, or close to zero (between -0.2 and 0.2) in the 3T data (compare Fig 2b and Extended Data Fig 8). Supplementary Fig 1 shows predictions from a single connection to illustrate behavioural variability was greater in the 3T data.
Controlling for parcellation and dimensionality of behaviour
To show that parcellating the amygdala yielded improvements in prediction accuracy, we also repeated the regression procedure with only the connections from our ROIs to the entire amygdala instead of all individual nuclei (a total of 28 possible predictors). All figure panels related to connections with the whole amygdala, instead of its seven distinct nuclei, were generated using identical methods (Fig 6 and Extended Data Fig 9; Supplementary Methods).
A second control analysis focused on whether it was advantageous to use behavioural scores derived from a factor analysis. The factorization of questionnaire scores was performed because we hypothesized functional connectivity in specific amygdala networks more likely relates to simpler functionally meaningful behavioural dimensions, as opposed to complex markers such as e.g., a person’s overall depression score. To validate this assumption, we repeated our analyses on participants’ total DSM and ASR Anxiety-Depression scores (ASR_Anxd_Raw). We used the same analysis pipeline and compared predictions achieved for the two depression markers to those obtained for our original four behavioural dimensions (Fig7 and Extended Data Fig 10).
Extended Data
Extended Data Fig. 1. Additional preprocessing to account for physiological noise.
a, Physiological noise correction steps: The minimally preprocessed HCP data was additionally corrected for physiological noise to improve the signal in temporal lobe and brainstem regions, the key areas for this study. All other data clean-up steps usually applied to generate fully preprocessed HCP data, specifically fix-denoising and motion correction, were applied at the same time. b, Signal to noise improvements: Illustration of the signal-to-noise improvements gained from this additional preprocessing step compared to standard full HCP preprocessing (in a subset of 100 participants). Top: Mean temporal signal to noise ratio (tSNR) obtained following our preprocessing pipeline; Bottom: Difference in tSNR between the preprocessing with and without physiological noise correction. The ratio of tSNRs (physio - noPhysio) / (physio + noPhysio) is illustrated. This shows tSNR gains in medial temporal lobe and medial prefrontal cortex but particularly subcortical and brainstem structures. c, tSNR improvements relative to no physiological noise correction in several ROIs for (a) respiratory, (b) respiratory + respiratory volume, (c) cardiac, (d) all three (PNM): The mean tSNR difference achieved with subsets of the physiological noise regressors is shown compared to the baseline of not performing any physiological noise correction. Improvements are illustrated for several regions of interest (ROIs) including amygdala, dorsal raphe (RN_DR), locus coeruleus (LC), and areas 25 and d32 in medial PFC. The regressors used were either just respiration, both respiration and respiratory volume over time (RVT), just cardiac activity, or all of the above (which is what was ultimately used in the main analysis). This shows that subcortical ROIs benefited more from physiological noise correction, with greatest improvements in brainstem nuclei, and that respiratory and cardiac regressors contributed about equally to the improvement, with the greatest improvements achieved when including all noise regressors. Error bars denote SEM; n=19 participants (individual datapoints shown).
Extended Data Fig. 2. Replication of average amygdala functional connectivity.
Average amygdala connectivity: The group connectome shown for the original n=200 3T young-adult HCP participants presented in Fig 1a was replicated in two other cohorts containing n=200 non-overlapping 3T-HCP participants and n=98 non-overlapping 7T-HCP participants. In the 3T data, we used an improved pre-processing pipeline to correct for physiological noise, as before. This was not possible in the 7T data where physiological noise regressors were not available. However, the 7T resting-state data has improved signal-to-noise in subcortical regions due to the higher field strength. Scale bar denotes Pearson’s correlation coefficient (functional connectivity), corrected for global absolute connectivity strength.
Extended Data Fig. 3. Amygdala parcellation at different clustering depths.
a, Generation of group average dense connectome for amygdala parcellation: Summary of the additional processing steps required to compute a group average connectome from the 200 individual concatenated restingstate fMRI (rs-fMRI) time-series. The group connectome, restricted to connectivity between amygdala voxels and the whole brain, formed the basis for the amygdala parcellation. b, Amygdala parcellation step by step: Individual steps of the hierarchical clustering algorithm led to increasing subdivisions of the amygdala. All steps leading up to our final parcellation (depth 12), and a few additional clustering steps beyond it (up to depth 15), are shown. Hierarchical clustering was performed on absolute functional connectivity values. Note, for example, the central nuclei splitting off in step 9 (left) and 12 (right). The 12 cluster solution had five unique clusters in each hemisphere and two connected clusters (same color = same cluster). For subsequent analyses, the corresponding clusters in each hemisphere were joined, resulting in a total of seven clusters.
Extended Data Fig. 4. Replication of amygdala parcellation.
a, For comparison, the parcellation of the amygdala obtained in the original n=200 3T participants is shown for the n=200 3T replication sample and the n=98 (all nonoverlapping) 7T participants (compare Fig 1b). This shows that the key subdivisions of the amygdala were replicated in these two additional parcellations. b, Visualization of the cluster centroids from a coronal (left) and sagittal (right) point of view illustrates the similarity of the three parcellations (diamond: original 3T parcellation; square: 3T replication; circle: 7T replication). c, Similarity of parcellations compared to null with contiguous symmetrical clusters: To quantify the similarity between the parcellations, two metrics are reported: the mean distance of the centroids and the % of overlapping voxels (i.e., voxels with identical labels). Null distributions respect the size and symmetry of the original parcellations but shuffle the location of the nuclei in a way that yields contiguous but non-overlapping clusters. This shows that the parcellations (left: comparison with 3T replication; right: with 7T replication) are more similar than expected by chance (onesided p-values from nonparametric test using permutation null distribution). Importantly, however, throughout the manuscript, we use the original 3T parcellation across all analyses. In addition, the choice of parcellation (which is based on the mean group connectivity) is orthogonal to the key findings reported in the manuscript which relate to interindividual variation that is ignored when generating the parcellation.
Extended Data Fig. 5. Replication of amygdala nuclei mean functional connectivity.
Strength of functional coupling (group average): The average amygdala nuclei to ROI functional connectivity, in all cases extracted based on the amygdala nuclei from the original n=200 3T parcellation, replicates across cohorts (top: original, bottom left: replication 3T, bottom right: replication 7T; compare Fig 2b), as confirmed in the strong correlation between these patterns (top right). Scale bar denotes Pearson’s r corrected for global absolute mean connectivity.
Extended Data Fig. 6. Distribution of behavioural scores and extracted latent behaviours.
a, Distribution of all behavioural markers included in the factor analysis shown in Fig 3 across the 200 HCP participants. For a full description of each score see Table 1 and Methods. b, Distribution of the latent behaviours generated from the factor analysis.
Extended Data Fig. 7. Replication of factor analysis.
Factor loadings show behavioural factor analysis replicability: The factor analysis computed to generate mental health dimensions in our original n=200 3T participants (left) replicated in all n=1206 HCP participants (2nd column) and the full set of n=400 3T and n=98 7T participants used in this manuscript (3rd and 4th column). Correlation coefficients and p-values refer to the similarity (two-sided t-test) with the original pattern shown on the left.
Extended Data Fig. 8. Detailed description of contributing anatomical networks.
a, Baseline average connectivity: Edges where functional connectivity was on average negative (<0.2, left), modulatory/zero (middle), or positive (>0.2, right; see also Fig 2b and Extended Data Fig 3b) in the group of all 3T participants are shown to aid interpretation of fingerprints. b, Anatomical fingerprint for smallest network at p<0.1 (left), p<0.05 (middle), peak (right): In addition to the smallest network of connections that reached significant out-ofsample predictions (using 3T-regression weights to predict 7T behavioural dimensions) shown in Fig 5c, here we show the smallest trend-wise significant (p<0.1) and significant (p<0.05 as in Fig 5c) network as well as the network associated with the peak prediction (compare Fig 5a); for precise p-value calculation in each case, see Methods and Results. All conventions are as in Fig 5: anatomical fingerprints show ROIs on the circumference (dark=subcortical), amygdala nuclei in the centre (lines are colour-coded); line width denotes the size of the absolute 3T regression coefficient; line style denotes its sign (continuous=positive; dashed=negative). In addition, here, a surrounding line is black if baseline functional connectivity in this edge is positive and grey if it is negative (defined as in a). Integer number indicates the number of edges shown; scatterplots underneath fingerprints show the associated out-of-sample 7T prediction.
Extended Data Fig. 9. Whole versus amygdala nuclei predictions.
a, For comparison, the out-of-sample prediction achieved using increasing numbers of edges for all behaviours and the nuclei-version shown in Fig 5a (here grey) is shown next to the out-of-sample predictions achieved using increasing numbers of whole-amygdala edges (coloured) which despite containing the same voxels in total performs worse than the nuclei version for all four mental health dimensions. b, Anatomical fingerprints associated with the peak out-of-sample prediction possible using wholeamygdala functional connectivity in a.
Extended Data Fig. 10. Amygdala functional connectivity relates better to dimensional behaviours than DSM scores.
a, b Predicting depression scores instead of dimensional markers: Out-of-sample prediction of 7T participant’s DSM scores, similar to ASR scores shown in Fig 7, is (a) not significant and (b) less accurate than three out of four of our dimensional behaviours (one-sided correlation to assess positive relationship). c, Overall predictions are worse for DSM compared to dimensional scores when using smaller sets of edges (bars shown in Fig 5a for dimensional behaviours are overlaid as coloured lines for comparison); plotting conventions as in Fig 5. d, Anatomical network at peak prediction: fingerprint shows the network of edges contributing to the prediction of DSM scores at the peak prediction (5 edges). e, Anatomical network at peak prediction: Similarly, anatomical fingerprint shows the network of edges contributing to the prediction of ASR scores at the peak prediction (35 edges, still only trend-wise significant) and its associated prediction; onesided correlation to assess positive relationships.
Supplementary Material
Acknowledgements
Data were provided by the Human Connectome Project, WU-Minn Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research; and by the McDonnell Center for Systems Neuroscience at Washington University. MCKF was funded by a Sir Henry Wellcome and Henry Dale Fellowship (103184/Z/13/Z – MKF; and 223263/Z/21/Z - MKF), MFSR was funded by an MRC grant (MR/P024955/1 - MFSR) and a Wellcome Senior Investigator Award (WT100973AIA - MFSR). The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.
Footnotes
Author contributions
MCKF and MFSR designed the study, MCKF, DEAJ and MFSR conceived analyses, MCKF and DEAJ wrote analysis code, LV, YT and SS gave analysis advice, LP helped with data pre-processing, and all authors wrote the manuscript.
Competing Interests statement
The authors declare no competing interests.
Data availability
All data used in the present study are available for download from the Human Connectome Project (www.humanconnectome.org). Users must apply for access and agree to the HCP data use terms (for details see https://www.humanconnectome.org/study/hcp-young-adult/data-use-terms). Here we used both Open Access and Restricted data. The masks of all ROIs used in this study as well as all individual amygdala nuclei generated here are available in the OSF repository https://doi.org/10.17605/OSF.IO/EGM2R.
Code availability statement
Code that allows users with HCP data access to replicate analyses is provided in the OSF repository https://doi.org/10.17605/OSF.IO/EGM2R. This includes code to plot all figures presented in the manuscript. Intermediate analysis outputs can be made available to registered HCP users. Please see the README file in the Scripts folder for further detail.
References
- 1.Bijsterbosch JD, et al. Stratification of MDD and GAD patients by resting state brain connectivity predicts cognitive bias. NeuroImage Clin. 2018;19:425–433. doi: 10.1016/j.nicl.2018.04.033. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Reggente N, et al. Multivariate resting-state functional connectivity predicts response to cognitive behavioral therapy in obsessive–compulsive disorder. Proc Natl Acad Sci. 2018;115:2222–2227. doi: 10.1073/pnas.1716686115. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Takagi Y, et al. A Neural Marker of Obsessive-Compulsive Disorder from Whole-Brain Functional Connectivity. Sci Rep. 2017;7:1–10. doi: 10.1038/s41598-017-07792-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Zeng L-L, et al. Identifying major depression using whole-brain functional connectivity: a multivariate pattern analysis. Brain J Neurol. 2012;135:1498–1507. doi: 10.1093/brain/aws059. [DOI] [PubMed] [Google Scholar]
- 5.Aggleton JP. The contribution of the amygdala to normal and abnormal emotional states. Trends Neurosci. 1993;16:328–333. doi: 10.1016/0166-2236(93)90110-8. [DOI] [PubMed] [Google Scholar]
- 6.Drevets WC. Neuroimaging abnormalities in the amygdala in mood disorders. Ann N Y Acad Sci. 2003;985:420–444. doi: 10.1111/j.1749-6632.2003.tb07098.x. [DOI] [PubMed] [Google Scholar]
- 7.Kennedy DP, Adolphs R. The social brain in psychiatric and neurological disorders. Trends Cogn Sci. 2012;16:559–572. doi: 10.1016/j.tics.2012.09.006. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Murray EA, Wise SP, Drevets WC. Localization of dysfunction in major depressive disorder: prefrontal cortex and amygdala. Biol Psychiatry. 2011;69:e43–54. doi: 10.1016/j.biopsych.2010.09.041. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Phelps EA, LeDoux JE. Contributions of the amygdala to emotion processing: from animal models to human behavior. Neuron. 2005;48:175–187. doi: 10.1016/j.neuron.2005.09.025. [DOI] [PubMed] [Google Scholar]
- 10.Adolphs R. The biology of fear. Curr Biol CB. 2013;23:R79–93. doi: 10.1016/j.cub.2012.11.055. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Bernardi S, Salzman CD. The contribution of nonhuman primate research to the understanding of emotion and cognition and its clinical relevance. Proc Natl Acad Sci. 2019;116:26305–26312. doi: 10.1073/pnas.1902293116. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Dal Monte O, Costa VD, Noble PL, Murray EA, Averbeck BB. Amygdala lesions in rhesus macaques decrease attention to threat. Nat Commun. 2015;6:10161. doi: 10.1038/ncomms10161. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Drevets WC, et al. Subgenual prefrontal cortex abnormalities in mood disorders. Nature. 1997;386:824–827. doi: 10.1038/386824a0. [DOI] [PubMed] [Google Scholar]
- 14.Chang SWC, et al. Neural mechanisms of social decision-making in the primate amygdala. Proc Natl Acad Sci U S A. 2015;112:16012–16017. doi: 10.1073/pnas.1514761112. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Murray EA, Wise SP. Interactions between orbital prefrontal cortex and amygdala: advanced cognition, learned responses and instinctive behaviors. Curr Opin Neurobiol. 2010;20:212–220. doi: 10.1016/j.conb.2010.02.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Amaral DG, Price JL, Pitkanen A, Carmichael ST. The Amygdala: Neurobiological Aspects of Emotion, Memory, and Mental Dysfunction. New York: Wiley-Liss; 1992. Anatomical organization of the primate amygdaloid complex; pp. 1–66. [Google Scholar]
- 17.Zangemeister L, Grabenhorst F, Schultz W. Neural Basis for Economic Saving Strategies in Human Amygdala-Prefrontal Reward Circuits. Curr Biol CB. 2016;26:3004–3013. doi: 10.1016/j.cub.2016.09.016. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Rudebeck PH, Rich EL, Mayberg HS. From bed to bench side: Reverse translation to optimize neuromodulation for mood disorders. Proc Natl Acad Sci. 2019;116:26288–26296. doi: 10.1073/pnas.1902287116. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Van Essen DC, et al. The WU-Minn Human Connectome Project: an overview. NeuroImage. 2013;80:62–79. doi: 10.1016/j.neuroimage.2013.05.041. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Insel T, et al. Research domain criteria (RDoC): toward a new classification framework for research on mental disorders. Am J Psychiatry. 2010;167:748–751. doi: 10.1176/appi.ajp.2010.09091379. [DOI] [PubMed] [Google Scholar]
- 21.Haro JM, et al. ROAMER: roadmap for mental health research in Europe. Int J Methods Psychiatr Res. 2014;23(Suppl 1):1–14. doi: 10.1002/mpr.1406. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Abivardi A, Bach DR. Deconstructing white matter connectivity of human amygdala nuclei with thalamus and cortex subdivisions in vivo. Hum Brain Mapp. 2017;38:3927–3940. doi: 10.1002/hbm.23639. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Bzdok D, Laird AR, Zilles K, Fox PT, Eickhoff SB. An investigation of the structural, connectional, and functional subspecialization in the human amygdala. Hum Brain Mapp. 2013;34:3247–3266. doi: 10.1002/hbm.22138. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Solano-Castiella E, et al. Parcellation of human amygdala in vivo using ultra high field structural MRI. NeuroImage. 2011;58:741–748. doi: 10.1016/j.neuroimage.2011.06.047. [DOI] [PubMed] [Google Scholar]
- 25.Smith SM, et al. Resting-state fMRI in the Human Connectome Project. NeuroImage. 2013;80:144–168. doi: 10.1016/j.neuroimage.2013.05.039. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Fudge JL, Kunishio K, Walsh P, Richard C, Haber SN. Amygdaloid projections to ventromedial striatal subterritories in the primate. Neuroscience. 2002;110:257–275. doi: 10.1016/s0306-4522(01)00546-2. [DOI] [PubMed] [Google Scholar]
- 27.Zikopoulos B, John YJ, García-Cabezas MÁ, Bunce JG, Barbas H. The intercalated nuclear complex of the primate amygdala. Neuroscience. 2016;330:267–290. doi: 10.1016/j.neuroscience.2016.05.052. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Mai JK, Majtanik M, Paxinos G. Atlas of the Human Brain. Academic Press; 2015. [Google Scholar]
- 29.Saygin ZM, et al. High-resolution magnetic resonance imaging reveals nuclei of the human amygdala: manual segmentation to automatic atlas. NeuroImage. 2017;155:370–382. doi: 10.1016/j.neuroimage.2017.04.046. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Ding S-L, et al. Comprehensive cellular-resolution atlas of the adult human brain. J Comp Neurol. 2016;524:3127–3481. doi: 10.1002/cne.24080. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Averbeck BB, Costa VD. Motivational neural circuits underlying reinforcement learning. Nat Neurosci. 2017;20:505–512. doi: 10.1038/nn.4506. [DOI] [PubMed] [Google Scholar]
- 32.Keuken MC, et al. Quantifying inter-individual anatomical variability in the subcortex using 7 T structural MRI. NeuroImage. 2014;94:40–46. doi: 10.1016/j.neuroimage.2014.03.032. [DOI] [PubMed] [Google Scholar]
- 33.Faull OK, Jenkinson M, Ezra M, Pattinson KT. Conditioned respiratory threat in the subdivisions of the human periaqueductal gray. eLife. 2016;5 doi: 10.7554/eLife.12047. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Betts MJ, Cardenas-Blanco A, Kanowski M, Jessen F, Düzel E. In vivo MRI assessment of the human locus coeruleus along its rostrocaudal extent in young and older adults. NeuroImage. 2017;163:150–159. doi: 10.1016/j.neuroimage.2017.09.042. [DOI] [PubMed] [Google Scholar]
- 35.Folloni D, et al. Dichotomous organization of amygdala/temporal-prefrontal bundles in both humans and monkeys. eLife. 2019;8:e47175. doi: 10.7554/eLife.47175. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Neubert F-X, Mars RB, Sallet J, Rushworth MFS. Connectivity reveals relationship of brain areas for reward-guided learning and decision making in human and monkey frontal cortex. Proc Natl Acad Sci U S A. 2015;112:E2695–2704. doi: 10.1073/pnas.1410767112. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Aggleton JP, Burton MJ, Passingham RE. Cortical and subcortical afferents to the amygdala of the rhesus monkey (Macaca mulatta) Brain Res. 1980;190:347–368. doi: 10.1016/0006-8993(80)90279-6. [DOI] [PubMed] [Google Scholar]
- 38.Ironside M, et al. Effect of Prefrontal Cortex Stimulation on Regulation of Amygdala Response to Threat in Individuals With Trait Anxiety: A Randomized Clinical Trial. JAMA Psychiatry. 2018 doi: 10.1001/jamapsychiatry.2018.2172. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Glasser MF, et al. A multi-modal parcellation of human cerebral cortex. Nature. 2016;536:171–178. doi: 10.1038/nature18933. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.O’Reilly JX, et al. Causal effect of disconnection lesions on interhemispheric functional connectivity in rhesus monkeys. Proc Natl Acad Sci U S A. 2013;110:13982–13987. doi: 10.1073/pnas.1305062110. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 41.Carmichael ST, Price JL. Sensory and premotor connections of the orbital and medial prefrontal cortex of macaque monkeys. J Comp Neurol. 1995;363:642–664. doi: 10.1002/cne.903630409. [DOI] [PubMed] [Google Scholar]
- 42.Freedman LJ, Insel TR, Smith Y. Subcortical projections of area 25 (subgenual cortex) of the macaque monkey. J Comp Neurol. 2000;421:172–188. [PubMed] [Google Scholar]
- 43.Sharma KK, Kelly EA, Pfeifer CW, Fudge JL. Translating Fear Circuitry: Amygdala Projections to Subgenual and Perigenual Anterior Cingulate in the Macaque. Cereb Cortex. 2020;30:550–562. doi: 10.1093/cercor/bhz106. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44.Harrison OK, Guell X, Klein-Flügge MC, Barry RL. Structural and resting state functional connectivity beyond the cortex. NeuroImage. 2021;240:118379. doi: 10.1016/j.neuroimage.2021.118379. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 45.Drysdale AT, et al. Resting-state connectivity biomarkers define neurophysiological subtypes of depression. Nat Med. 2017;23:28–38. doi: 10.1038/nm.4246. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 46.Gillan CM, Kosinski M, Whelan R, Phelps EA, Daw ND. Characterizing a psychiatric symptom dimension related to deficits in goal-directed control. eLife. 2016;5:e11305. doi: 10.7554/eLife.11305. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 47.Rouault M, Seow T, Gillan CM, Fleming SM. Psychiatric Symptom Dimensions Are Associated With Dissociable Shifts in Metacognition but Not Task Performance. Biol Psychiatry. 2018;84:443–451. doi: 10.1016/j.biopsych.2017.12.017. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 48.Williams LM. Defining biotypes for depression and anxiety based on large-scale circuit dysfunction: a theoretical review of the evidence and future directions for clinical translation. Depress Anxiety. 2017;34:9–24. doi: 10.1002/da.22556. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 49.Holtzheimer PE, et al. Subcallosal cingulate deep brain stimulation for treatment-resistant depression: a multisite, randomised, sham-controlled trial. Lancet Psychiatry. 2017;4:839–849. doi: 10.1016/S2215-0366(17)30371-1. [DOI] [PubMed] [Google Scholar]
- 50.Dandekar MP, Fenoy AJ, Carvalho AF, Soares JC, Quevedo J. Deep brain stimulation for treatment-resistant depression: an integrative review of preclinical and clinical findings and translational implications. Mol Psychiatry. 2018;23:1094–1112. doi: 10.1038/mp.2018.2. [DOI] [PubMed] [Google Scholar]
- 51.Riva-Posse P, et al. A connectomic approach for subcallosal cingulate deep brain stimulation surgery: prospective targeting in treatment-resistant depression. Mol Psychiatry. 2018;23:843–849. doi: 10.1038/mp.2017.59. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 52.Johansen-Berg H, et al. Anatomical connectivity of the subgenual cingulate region targeted with deep brain stimulation for treatment-resistant depression. Cereb Cortex N Y N 1991. 2008;18:1374–1383. doi: 10.1093/cercor/bhm167. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 53.Dichter GS, Gibbs D, Smoski MJ. A systematic review of relations between resting-state functional-MRI and treatment response in major depressive disorder. J Affect Disord. 2015;172:8–17. doi: 10.1016/j.jad.2014.09.028. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 54.Salomons TV, et al. Resting-state cortico-thalamic-striatal connectivity predicts response to dorsomedial prefrontal rTMS in major depressive disorder. Neuropsychopharmacol Off Publ Am Coll Neuropsychopharmacol. 2014;39:488–498. doi: 10.1038/npp.2013.222. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 55.Furtado CP, et al. An investigation of medial temporal lobe changes and cognition following antidepressant response: a prospective rTMS study. Brain Stimulat. 2013;6:346. doi: 10.1016/j.brs.2012.06.006. [DOI] [PubMed] [Google Scholar]
- 56.Fox MD, Buckner RL, White MP, Greicius MD, Pascual-Leone A. Efficacy of transcranial magnetic stimulation targets for depression is related to intrinsic functional connectivity with the subgenual cingulate. Biol Psychiatry. 2012;72:595–603. doi: 10.1016/j.biopsych.2012.04.028. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 57.Verhagen L, et al. Offline impact of transcranial focused ultrasound on cortical activation in primates. eLife. 2019;8 doi: 10.7554/eLife.40541. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 58.Bongioanni A, et al. Activation and disruption of a neural mechanism for novel choice in monkeys. Nature. 2021:1–5. doi: 10.1038/s41586-020-03115-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 59.Blackmore J, Shrivastava S, Sallet J, Butler CR, Cleveland RO. Ultrasound Neuromodulation: A Review of Results, Mechanisms and Safety. Ultrasound Med Biol. 2019;45:1509–1536. doi: 10.1016/j.ultrasmedbio.2018.12.015. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 60.Folloni D, et al. Manipulation of Subcortical and Deep Cortical Activity in the Primate Brain Using Transcranial Focused Ultrasound Stimulation. Neuron. 2019;101:1109–1116.:e5. doi: 10.1016/j.neuron.2019.01.019. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 61.Browning M, Behrens TE, Jocham G, O’Reilly JX, Bishop SJ. Anxious individuals have difficulty learning the causal statistics of aversive environments. Nat Neurosci. 2015;18:590–596. doi: 10.1038/nn.3961. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 62.Scholl J, Klein-Flügge M. Understanding psychiatric disorder by capturing ecologically relevant features of learning and decision-making. Behav Brain Res. 2018;355:56–75. doi: 10.1016/j.bbr.2017.09.050. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 63.Klein-Flügge MC, Bongioanni A, Rushworth MFS. Medial and orbital frontal cortex in decision-making and flexible behavior. Neuron. 2022:S0896-6273(22)00463–9. doi: 10.1016/j.neuron.2022.05.022. [DOI] [PubMed] [Google Scholar]
- 64.Kolling N, Scholl J, Chekroud A, Trier HA, Rushworth MFS. Prospection, Perseverance, and Insight in Sequential Behavior. Neuron. 2018;99:1069–1082.:e7. doi: 10.1016/j.neuron.2018.08.018. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 65.Beliveau V, et al. A High-Resolution In Vivo Atlas of the Human Brain’s Serotonin System. J Neurosci Off J Soc Neurosci. 2017;37:120–128. doi: 10.1523/JNEUROSCI.2830-16.2016. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 66.Saulin A, Savli M, Lanzenberger R. Serotonin and molecular neuroimaging in humans using PET. Amino Acids. 2012;42:2039–2057. doi: 10.1007/s00726-011-1078-9. [DOI] [PubMed] [Google Scholar]
- 67.Palomero-Gallagher N, Amunts K, Zilles K. Transmitter Receptor Distribution in the Human Brain. 2015 doi: 10.1016/B978-0-12-397025-1.00221-9. [DOI] [Google Scholar]
- 68.van den Brink RL, Pfeffer T, Donner TH. Brainstem Modulation of Large-Scale Intrinsic Cortical Activity Correlations. Front Hum Neurosci. 2019;13:340. doi: 10.3389/fnhum.2019.00340. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 69.Zilles K, et al. Architectonics of the human cerebral cortex and transmitter receptor fingerprints: reconciling functional neuroanatomy and neurochemistry. Eur Neuropsychopharmacol. 2002;12:587–599. doi: 10.1016/s0924-977x(02)00108-6. [DOI] [PubMed] [Google Scholar]
- 70.Almeida JRC, et al. Reduced gray matter volume in ventral prefrontal cortex but not amygdala in bipolar disorder: significant effects of gender and trait anxiety. Psychiatry Res. 2009;171:54–68. doi: 10.1016/j.pscychresns.2008.02.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 71.Klein-Flügge MC, Wittmann MK, Shpektor A, Jensen DEA, Rushworth MFS. Multiple associative structures created by reinforcement and incidental statistical learning mechanisms. Nat Commun. 2019;10:1–15. doi: 10.1038/s41467-019-12557-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 72.Wittmann MK, et al. Global reward state affects learning and activity in raphe nucleus and anterior insula in monkeys. Nat Commun. 2020;11:3771. doi: 10.1038/s41467-020-17343-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 73.Schildkraut JJ. The catecholamine hypothesis of affective disorders: a review of supporting evidence. Am J Psychiatry. 1965;122:509–522. doi: 10.1176/ajp.122.5.509. [DOI] [PubMed] [Google Scholar]
- 74.Oishi Y, et al. Slow-wave sleep is controlled by a subset of nucleus accumbens core neurons in mice. Nat Commun. 2017;8:1–12. doi: 10.1038/s41467-017-00781-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 75.Uğurbil K, et al. Pushing spatial and temporal resolution for functional and diffusion MRI in the Human Connectome Project. NeuroImage. 2013;80:80–104. doi: 10.1016/j.neuroimage.2013.05.012. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 76.Glasser MF, et al. The minimal preprocessing pipelines for the Human Connectome Project. NeuroImage. 2013;80:105–124. doi: 10.1016/j.neuroimage.2013.04.127. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 77.Brooks JCW, et al. Physiological noise modelling for spinal functional magnetic resonance imaging studies. NeuroImage. 2008;39:680–692. doi: 10.1016/j.neuroimage.2007.09.018. [DOI] [PubMed] [Google Scholar]
- 78.Smith SM, Hyvärinen A, Varoquaux G, Miller KL, Beckmann CF. Group-PCA for very large fMRI datasets. Neuroimage. 2014;101:738–749. doi: 10.1016/j.neuroimage.2014.07.051. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 79.Gershon RC, et al. NIH Toolbox for Assessment of Neurological and Behavioral Function. Neurology. 2013;80:S2–S6. doi: 10.1212/WNL.0b013e3182872e5f. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 80.Salsman JM, et al. Emotion assessment using the NIH Toolbox. Neurology. 2013;80:S76–S86. doi: 10.1212/WNL.0b013e3182872e11. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 81.Buysse DJ, Reynolds CF, Monk TH, Berman SR, Kupfer DJ. The Pittsburgh Sleep Quality Index: a new instrument for psychiatric practice and research. Psychiatry Res. 1989;28:193–213. doi: 10.1016/0165-1781(89)90047-4. [DOI] [PubMed] [Google Scholar]
- 82.Heine SJ, Buchtel EE. Personality: the universal and the culturally specific. Annu Rev Psychol. 2009;60:369–394. doi: 10.1146/annurev.psych.60.110707.163655. [DOI] [PubMed] [Google Scholar]
- 83.McCrae RR, Costa PT., Jr A contemplated revision of the NEO Five-Factor Inventory. Personal Individ Differ. 2004;36:587–596. [Google Scholar]
- 84.Raîche G, Blais J, Montréal UD. Non Graphical Solutions for the Cattell’s Scree Test. Journal of Research Methods for the Behavioral and Social Sciences. 2013:23–29. [Google Scholar]
- 85.Smith SM, et al. A positive-negative mode of population covariation links brain connectivity, demographics and behavior. Nat Neurosci. 2015;18:1565–1567. doi: 10.1038/nn.4125. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
All data used in the present study are available for download from the Human Connectome Project (www.humanconnectome.org). Users must apply for access and agree to the HCP data use terms (for details see https://www.humanconnectome.org/study/hcp-young-adult/data-use-terms). Here we used both Open Access and Restricted data. The masks of all ROIs used in this study as well as all individual amygdala nuclei generated here are available in the OSF repository https://doi.org/10.17605/OSF.IO/EGM2R.
Code that allows users with HCP data access to replicate analyses is provided in the OSF repository https://doi.org/10.17605/OSF.IO/EGM2R. This includes code to plot all figures presented in the manuscript. Intermediate analysis outputs can be made available to registered HCP users. Please see the README file in the Scripts folder for further detail.